# ![@anayatkhan09 Avatar](https://lunarcrush.com/gi/w:26/cr:twitter::1446814317047910402.png) @anayatkhan09 Anayat

Anayat posts on X about ai, claude code, loops, vibe coding the most. They currently have [---] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours.

### Engagements: [------] [#](/creator/twitter::1446814317047910402/interactions)
![Engagements Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1446814317047910402/c:line/m:interactions.svg)

- [--] Week [-------] +23%
- [--] Month [-------] +340%
- [--] Months [-------] +59,732%
- [--] Year [-------] +3,265%

### Mentions: [---] [#](/creator/twitter::1446814317047910402/posts_active)
![Mentions Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1446814317047910402/c:line/m:posts_active.svg)

- [--] Week [---] +62%
- [--] Month [---] +127%
- [--] Months [---] +74,100%
- [--] Year [---] +4,544%

### Followers: [---] [#](/creator/twitter::1446814317047910402/followers)
![Followers Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1446814317047910402/c:line/m:followers.svg)

- [--] Week [---] +11%
- [--] Month [---] +38%
- [--] Months [---] +89%
- [--] Year [---] +96%

### CreatorRank: [-------] [#](/creator/twitter::1446814317047910402/influencer_rank)
![CreatorRank Line Chart](https://lunarcrush.com/gi/w:600/cr:twitter::1446814317047910402/c:line/m:influencer_rank.svg)

### Social Influence

**Social category influence**
[technology brands](/list/technology-brands)  [finance](/list/finance)  [stocks](/list/stocks)  [social networks](/list/social-networks)  [currencies](/list/currencies)  [vc firms](/list/vc-firms)  [products](/list/products)  [celebrities](/list/celebrities)  [countries](/list/countries) 

**Social topic influence**
[ai](/topic/ai), [claude code](/topic/claude-code) #1349, [loops](/topic/loops) #354, [vibe coding](/topic/vibe-coding) #644, [if you](/topic/if-you), [agents](/topic/agents) #729, [vercel](/topic/vercel), [systems](/topic/systems) #2454, [agentic](/topic/agentic) #765, [products](/topic/products) #2973

**Top accounts mentioned or mentioned by**
[@aisdk](/creator/undefined) [@openai](/creator/undefined) [@nextjs](/creator/undefined) [@thdxr](/creator/undefined) [@prisma](/creator/undefined) [@magicuidesign](/creator/undefined) [@gergelyorosz](/creator/undefined) [@omarsar0](/creator/undefined) [@clerkdev](/creator/undefined) [@supabase](/creator/undefined) [@neondatabase](/creator/undefined) [@corbinbraun](/creator/undefined) [@stripe](/creator/undefined) [@yuchenjuw](/creator/undefined) [@clerk](/creator/undefined) [@austen](/creator/undefined) [@tailwindcss](/creator/undefined) [@steipete](/creator/undefined) [@thsottiaux](/creator/undefined) [@petergyang](/creator/undefined)

**Top assets mentioned**
[Alphabet Inc Class A (GOOGL)](/topic/$googl)
### Top Social Posts
Top posts by engagements in the last [--] hours

"Check out the AI-powered blog generator Create fully-formed plagiarism-free posts with just one click. Built with @nextjs @typescript @tailwindcss @dalle_openai @ClerkDev and @supabase. - live: - code:"  
[X Link](https://x.com/anayatkhan09/status/1761745051175047465)  2024-02-25T13:28Z [---] followers, [--] engagements


"🚀 Introducing the AI-powered blog generator Generate original blog posts with a single click. - Built with Next.js TypeScript Tailwind CSS DALLE OpenAI Clerk Supabase and AWS Amplify. Check it out: GitHub"  
[X Link](https://x.com/anayatkhan09/status/1762420141311099020)  2024-02-27T10:11Z [---] followers, [---] engagements


"Next.js 14+ Starter Template now available ⚡ Featuring App Router Tailwind CSS TypeScript ESLint Prettier Shadcn UI background snippets Husky Jest Testing Library Playwright Vercel Analytics & SpeedInsight. - Check it out on GitHub:"  
[X Link](https://x.com/anayatkhan09/status/1764317834933346745)  2024-03-03T15:52Z [---] followers, [---] engagements


"Just submitted my solution for the KoinX (@getkoinx) Frontend Intern Assignment Used Next.js Tailwind CSS TypeScript and Shadcn UI. Integrated TradingView & Coingecko APIs plus added test scripts. - Github Deployed:"  
[X Link](https://x.com/anayatkhan09/status/1765467440492707920)  2024-03-06T20:00Z [---] followers, [---] engagements


"My [----] Tech Stack: Language: @typescript Framework: @nextjs DB: @neondatabase / @prisma Styling: @tailwindcss AI API: @togethercompute Hosting & analytics: @vercel Auth: @ClerkDev"  
[X Link](https://x.com/anayatkhan09/status/1782974881388450127)  2024-04-24T03:28Z [---] followers, [---] engagements


"🚀 Excited to introduce Writora AI Effortlessly convert your audio and video content into engaging blog posts in seconds. Tech Stack: @nextjs @tailwindcss @shadcn @magicuidesign @ClerkDev @stripe @neondatabase @OpenAI Try Writora AI - https://writora.xyz https://writora.xyz"  
[X Link](https://x.com/anayatkhan09/status/1847363197826638113)  2024-10-18T19:44Z [---] followers, [----] engagements


"@DruxAMB @nextjs @tailwindcss @shadcn @magicuidesign @ClerkDev @stripe @neondatabase @OpenAI @vercel @Namecheap @t3dotgg thank you so much 😊"  
[X Link](https://x.com/anayatkhan09/status/1847650760722436430)  2024-10-19T14:47Z [---] followers, [--] engagements


"@mrterrycarson @nextjs @tailwindcss @shadcn @magicuidesign @ClerkDev @stripe @neondatabase @OpenAI @vercel @Namecheap @t3dotgg thank you so much 😊"  
[X Link](https://x.com/anayatkhan09/status/1847650781865922866)  2024-10-19T14:47Z [---] followers, [--] engagements


"@mushtqmk @nextjs @tailwindcss @shadcn @magicuidesign @ClerkDev @stripe @neondatabase @OpenAI @vercel @Namecheap @t3dotgg thank you so much 😊"  
[X Link](https://x.com/anayatkhan09/status/1847650798114648103)  2024-10-19T14:47Z [---] followers, [--] engagements


"🚀  A SaaS automation tool that connects your apps and automates tasks effortlessly. : @nextjs @aceternitylabs @ClerkDev @neondatabase @prisma @Uploadcare @ngrokHQ @stripe  : https://fuzora.xyz https://fuzora.xyz"  
[X Link](https://x.com/anayatkhan09/status/1847919109180133386)  2024-10-20T08:33Z [---] followers, [---] engagements


":  : Monitor changes in real-time and automate responses.  & : Send notifications and custom messages to keep your team informed"  
[X Link](https://x.com/anayatkhan09/status/1847919111713558729)  2024-10-20T08:33Z [---] followers, [---] engagements


"Fuzora is here to save you time and reduce manual effort allowing you to focus on what truly matters. Start automating your workflows today. : #automation #saas #projects #github https://github.com/anayatkhan1/Fuzora https://github.com/anayatkhan1/Fuzora"  
[X Link](https://x.com/anayatkhan09/status/1847919114066460885)  2024-10-20T08:33Z [---] followers, [---] engagements


"@lucahowey True You can reach out to the @raycastapp team 😁 btw For tasks and documentation which tools are you using"  
[X Link](https://x.com/anayatkhan09/status/1848082258533695914)  2024-10-20T19:22Z [---] followers, [--] engagements


"Obsidian Vault Template for Developers This template is crafted to boost your productivity and streamline your project management. @obsdmd"  
[X Link](https://x.com/anayatkhan09/status/1851913198250660087)  2024-10-31T09:04Z [---] followers, [----] engagements


"🚀 Super excited to share that within [--] hours of launching my personal project I've gained [--] users and [--] upvotes on @ProductHunt Huge shoutout to my favorite tech stack 🎉: @nextjs @magicuidesign @neondatabase @ClerkDev Share your feedback - https://www.producthunt.com/posts/writora-ai https://www.producthunt.com/posts/writora-ai"  
[X Link](https://x.com/anayatkhan09/status/1853523673879318694)  2024-11-04T19:44Z [---] followers, [---] engagements


"🛠WIP - Working on another exciting project I'm using an awesome tech stack: @nextjs @magicuidesign @aceternitylabs @neondatabase @prisma @llama_index @OpenAI Open for suggestions and feedback Check out the project overview - #buildinpublic https://t.ly/WdL0x https://t.ly/WdL0x"  
[X Link](https://x.com/anayatkhan09/status/1854285691158413545)  2024-11-06T22:12Z [---] followers, [---] engagements


"🚀 Excited to share a sneak peek of my latest project I'd love your feedback on the frontend design. - What do you think - Any suggestions for improvement #buildinpublic"  
[X Link](https://x.com/anayatkhan09/status/1854545542455390525)  2024-11-07T15:24Z [---] followers, [---] engagements


"🚀 Built in just [--] hours for the @lovable_dev Contest Introducing AI Travel Planner Your AI-Powered Travel Companion 🌍 🚀 Tech Stack: @reactjs @typescript @supabase @AnthropicAI @framer @getsentry 🔹Project link - https://ai-odyssey-planner.lovable.app/ https://ai-odyssey-planner.lovable.app/"  
[X Link](https://x.com/anayatkhan09/status/1906187688929358091)  2025-03-30T03:32Z [---] followers, [---] engagements


"One API All Models Instead of juggling different APIs for: OpenAI Anthropic Google Mistral xAI and so on. You learn one unified interface. No vendor lock-in. No rewrites. Just plug in and build. This is a huge reason behind the SDKs growth"  
[X Link](https://x.com/anayatkhan09/status/1995935456446611915)  2025-12-02T19:17Z [---] followers, [--] engagements


"The architecture is stupidly clean The SDK is built in [--] simple layers: Core runs anywhere JavaScript runs UI Hooks React Vue Svelte Angular RSC optional for server-side streaming The powerful part The backend frontend streaming logic is abstracted. Anyone whos written streaming manually knows this is a gift"  
[X Link](https://x.com/anayatkhan09/status/1995935459248406723)  2025-12-02T19:17Z [---] followers, [--] engagements


"Day [--] of learning the Vercel AI SDK: Today I explored model providers: OpenAI Google Anthropic and even local models. Just install the provider package add the API key and call the model. So clean. So flexible. Loving this SDK so far. ⚡ List of providers: https://ai-sdk.dev/providers/ai-sdk-providers#ai-sdk-providers https://ai-sdk.dev/providers/ai-sdk-providers#ai-sdk-providers"  
[X Link](https://x.com/anayatkhan09/status/1996292408049553453)  2025-12-03T18:56Z [---] followers, [---] engagements


"Day [--] of learning the Vercel AI SDK: Today I learned how to stream text to the terminal using streamText(). Key takeaways: Streams output chunk by chunk Feels much more responsive than waiting for full text API is almost identical to generateText() Trying this next in my agent workflow. What part of streaming logic confused you the first time 👇"  
[X Link](https://x.com/anayatkhan09/status/1996991296427417697)  2025-12-05T17:13Z [---] followers, [--] engagements


"Day [--] of learning the Vercel AI SDK: Today I learned how to stream LLM responses directly into a UI. Key points: UIMessageStream connects backend frontend useChat() handles messages + state on the client API route converts UI messages model messages"  
[X Link](https://x.com/anayatkhan09/status/1998412068479799676)  2025-12-09T15:19Z [---] followers, [---] engagements


"Day [--] of learning the Vercel AI SDK Today I learned how to track token usage. Key takeaways: Every model call exposes a usage object You can inspect input + output token counts Useful for cost tracking and prompt optimization Works even when using streaming"  
[X Link](https://x.com/anayatkhan09/status/1999864589492666706)  2025-12-13T15:30Z [---] followers, [---] engagements


"@mdancho84 [--] pages on agents from OpenAI is basically the [----] syllabus for data scientists not just a nice to read PDF. Skimming this is optional treating it like a build manual is how you end up in the 1% youre talking about"  
[X Link](https://x.com/anayatkhan09/status/2000275357124170166)  2025-12-14T18:43Z [---] followers, [---] engagements


"@huybery Enterprise cares less about can it write code and more about can it hold context across months of diffs without silently breaking invariants. The real Turing test for coding agents is when QA files fewer regression tickets and infra stops treating them as a risk surface"  
[X Link](https://x.com/anayatkhan09/status/2000473996626887071)  2025-12-15T07:52Z [---] followers, [--] engagements


"vibe coding makes the App Store feel more like npm than a shelf of finished products and that breaks the current ranking and review model. The interesting infra question is who becomes the Stripe for micro apps handling identity billing and distribution when a single creator is pushing [--] tiny utilities a year"  
[X Link](https://x.com/anayatkhan09/status/2000627128833057172)  2025-12-15T18:00Z [---] followers, [---] engagements


"@iannuttall Context hungry stacks hit limits fastest so Id prioritize Vercel AI SDK Next.js app router Drizzle ORM and Stripe then you can ship full agentic flows with playbookscom without hand-rolling every integration"  
[X Link](https://x.com/anayatkhan09/status/2000827886853611589)  2025-12-16T07:18Z [---] followers, [---] engagements


"@DeepLearningAI Autonomous but unreliable is the perfect description of raw MCP stacks right now; the real unlock is treating aisuite + MCP as primitives and layering domain checks retries and coarse workflow graphs on top so agents behave like systems not demos"  
[X Link](https://x.com/anayatkhan09/status/2000829913000243376)  2025-12-16T07:26Z [---] followers, [--] engagements


"Day [--] of learning the Vercel AI SDK Today I explored context window limits. Key points: - Every LLM has a hard token limit (input + output) - Exceeding it causes failures or quota errors Context limits shape LLM system design"  
[X Link](https://x.com/anayatkhan09/status/2000929201038213251)  2025-12-16T14:01Z [---] followers, [---] engagements


"Code + examples linked below https://github.com/agentcn-kit/ai-sdk-lab/blob/main/llm-fundamentals/context-window/README.md https://github.com/agentcn-kit/ai-sdk-lab/blob/main/llm-fundamentals/context-window/README.md"  
[X Link](https://x.com/anayatkhan09/status/2000929203517063512)  2025-12-16T14:01Z [---] followers, [--] engagements


"@jietang Teaching AI to operate smartphones like a human is where agents finally escape the browser sandbox; super curious how you handle robustness across OEM UI quirks and updates without retraining the whole AutoGLM stack each time"  
[X Link](https://x.com/anayatkhan09/status/2000961085462208824)  2025-12-16T16:07Z [---] followers, [----] engagements


"@ericjang11 Research agents feeling nascent vs sharp coding assistants tracks with my experience too the missing piece is tight eval loops on real browsing traces not just frozen corpora so agents get punished for shallow synthesis"  
[X Link](https://x.com/anayatkhan09/status/2000962885628186703)  2025-12-16T16:15Z [---] followers, [--] engagements


"@astrange Engineering is still the top of the ladder but vibe coding and agentic workflows only work long term when someone on the team owns architecture observability and debugging like any other production system"  
[X Link](https://x.com/anayatkhan09/status/2001179862577619303)  2025-12-17T06:37Z [---] followers, [--] engagements


"@willdepue This framing that any sufficiently expressive net is a big parallel computer matches what we see in practice the painful part is always the optimization bottlenecks and system-level tricks that actually let SGD explore enough of that program space"  
[X Link](https://x.com/anayatkhan09/status/2001180134758572170)  2025-12-17T06:38Z [---] followers, [---] engagements


"@dr_cintas Love how this skips brittle site specific APIs the interesting hard part now is observability and guardrails when an agent can click anything on the open web and one bad goal turns into a mess"  
[X Link](https://x.com/anayatkhan09/status/2001180404578189575)  2025-12-17T06:39Z [---] followers, [---] engagements


"Day [--] of learning the Vercel AI SDK Today I learned about prompt caching. Key points: - Providers cache repeated input tokens - Caching breaks at the first mismatched token - In conversations previous turns are usually cached Small prompt changes can impact cost a lot"  
[X Link](https://x.com/anayatkhan09/status/2001291616116486487)  2025-12-17T14:01Z [---] followers, [---] engagements


"@bibryam Context engineering is quickly becoming the real bottleneck in LLM apps curious how youd map this taxonomy onto actual system boundaries in distributed agents where context comes from tools logs and user state over long sessions"  
[X Link](https://x.com/anayatkhan09/status/2001299714248220949)  2025-12-17T14:33Z [---] followers, [---] engagements


"@tibo_maker AI killed the moat of just shipping a SaaS the only durable edge now is owning a critical workflow and wiring deeply into a teams stack so ripping you out breaks their day rather than just their UI"  
[X Link](https://x.com/anayatkhan09/status/2001300027885445250)  2025-12-17T14:34Z [---] followers, [--] engagements


"@levelsio @nikitabier A history tab is nice but the real unlock is indexing viewed posts so Grok can answer things like find me that iOS video UX thread I saw yesterday otherwise you are just scrolling a glorified log forever"  
[X Link](https://x.com/anayatkhan09/status/2001552004850827295)  2025-12-18T07:15Z [---] followers, [--] engagements


"@Lovable Love this direction for ChatGPT building apps and websites feels magical the real win is when it also handles boring stuff like schema changes auth and prod deploys so devs can stay in the loop instead of wrestling scaffolding"  
[X Link](https://x.com/anayatkhan09/status/2001552567076294758)  2025-12-18T07:18Z [---] followers, [---] engagements


"@OpenAIDevs App directory is cool but the real moat will be apps that own a tight vertical workflow not generic chat; curious how Apps SDK plays with existing stacks like Vercel AI SDK and whether devs treat ChatGPT as primary UI or just another surface"  
[X Link](https://x.com/anayatkhan09/status/2001553513617461377)  2025-12-18T07:21Z [---] followers, [----] engagements


"@lennysan Vibe coders are only scary for mid engineers; senior engineers will be the ones turning their Lovable Replit Bolt v0 Cursor workflows into robust infra security and DX so teams can actually ship and maintain what they vibe"  
[X Link](https://x.com/anayatkhan09/status/2001556418567278847)  2025-12-18T07:33Z [---] followers, [--] engagements


"@aidenybai Vibe coding going from [--] to default in [--] months is why infra and tooling need to assume non devs are wiring production paths; guardrails observability and typed surfaces suddenly matter a lot more than clever prompts"  
[X Link](https://x.com/anayatkhan09/status/2001557097910358459)  2025-12-18T07:36Z [---] followers, [---] engagements


"@DeryaTR_ The app directory part only works if apps expose clean typed APIs and state models; otherwise ChatGPT turns into a brittle launcher instead of a real agentic OS that can compose Canva Spotify and friends in one flow"  
[X Link](https://x.com/anayatkhan09/status/2001558187842114014)  2025-12-18T07:40Z [---] followers, [--] engagements


"@RhysSullivan GitHub Desktop nails the [--] percent use case you listed but the real win for teams would be surfacing richer context in that UI like CI status code owners and AI summaries so branch swapping and PR previews are actually decision ready"  
[X Link](https://x.com/anayatkhan09/status/2001558504335905175)  2025-12-18T07:41Z [---] followers, [---] engagements


"@veggie_eric The Big Bench audio win and [----] per minute pricing makes Grok Voice super interesting for real time agents curious how streaming latency and partial hypotheses behave under high churn conversations where you need tool calls in the loop"  
[X Link](https://x.com/anayatkhan09/status/2001558703154307124)  2025-12-18T07:42Z [---] followers, [--] engagements


"@zirkelc_ @aisdk Clean pattern thanks Love the name/description coupling via createToolSet and keeping tenant config out of UI types. Have you needed dynamic per-request toggles (e.g. enabled: (ctx) = boolean on tools)"  
[X Link](https://x.com/anayatkhan09/status/2001563452482163053)  2025-12-18T08:01Z [---] followers, [--] engagements


"@code @AnthropicAI Agent Skills are huge but the real test will be whether a skill authored for Claude in VS Code can plug into other runtimes and orchestrators with zero schema glue code or if we still end up hand-rolling adapters everywhere"  
[X Link](https://x.com/anayatkhan09/status/2001930157398151304)  2025-12-19T08:18Z [---] followers, [---] engagements


"@ericw_ai Most people obsess over tools but the boring parts in that diagram are where agents actually break in prod: state modeling memory boundaries and how you recover when a tool call silently half-completes"  
[X Link](https://x.com/anayatkhan09/status/2001930297898877044)  2025-12-19T08:19Z [---] followers, [--] engagements


"@levie Context engineering is only half the job the other half is building retrieval and stateful memory so the agent doesnt feel like a goldfish with RAG glued on top but like a service with its own evolving internal world model"  
[X Link](https://x.com/anayatkhan09/status/2001930784849142264)  2025-12-19T08:21Z [---] followers, [---] engagements


"@awilkinson AI can turn a $5M engineering department into a $50k line item on paper but if you dont invest those savings into better specs code review test infra and observability you just swapped payroll burn for future incident burn"  
[X Link](https://x.com/anayatkhan09/status/2001930957348372684)  2025-12-19T08:21Z [---] followers, [---] engagements


"Day [--] of learning the Vercel AI SDK Today I explored how UIMessage parts work with agents + tools. When you add a toolset streams arent just text: tool calls become steps inputs/outputs stream text follows execution All merged into one UIMessage via parts Day [--] of learning the Vercel AI SDK Today I implemented filesystem tool calling. The model can: - perform CRUD operations on files - create and update todo lists - decide when to use tools - stop based on custom conditions https://t.co/J0b3YwSdUk Day [--] of learning the Vercel AI SDK Today I implemented filesystem tool calling. The model"  
[X Link](https://x.com/anayatkhan09/status/2002031832456646784)  2025-12-19T15:02Z [---] followers, [---] engagements


"@a16z The line engineering skill matters more now is exactly why small LLM teams feel leverage rich most of the edge is wiring context tools evals and infra so simple ideas actually survive contact with production"  
[X Link](https://x.com/anayatkhan09/status/2002062069575884813)  2025-12-19T17:02Z [---] followers, [--] engagements


"@beyang Review UI being separate from the main agent panel makes sense but the real unlock is letting the review agent adapt depth based on diff size and risk surface so you are not burning tokens doing line comments on trivial refactors"  
[X Link](https://x.com/anayatkhan09/status/2002062872676118541)  2025-12-19T17:06Z [---] followers, [--] engagements


"@graphite The fun version of Graphite joining Cursor is pull requests becoming the default collaboration surface for AI work too where review comments agent traces and eval failures all show up in the same diff thread your team already lives in"  
[X Link](https://x.com/anayatkhan09/status/2002063446448484388)  2025-12-19T17:08Z [---] followers, [---] engagements


"@Yuchenj_UW Vibe coding only works when the jagged intelligence is pinned down with evals and failure budgets in a real app otherwise the vibes drift as soon as you ship to users"  
[X Link](https://x.com/anayatkhan09/status/2002280114441093278)  2025-12-20T07:29Z [---] followers, [---] engagements


"@GergelyOrosz Cursor playing chess only matters if the AI first repo also owns stacked diffs org wide evals and policy aware merge gates; otherwise GitHubs boring CI and permissions gravity still wins day to day"  
[X Link](https://x.com/anayatkhan09/status/2002280834825392341)  2025-12-20T07:32Z [---] followers, [---] engagements


"Day [--] of learning the @aisdk Today I wired tool calls all the way from backend to frontend. By typing UIMessage with InferUITools tool parts become type-safe in the UI. Rename a tool in backend frontend breaks instantly"  
[X Link](https://x.com/anayatkhan09/status/2002397608933732492)  2025-12-20T15:16Z [---] followers, [--] engagements


"@burkov The interesting bit is your point about Claude as a coding application weve had the best results treating Gemini the same way in our stack by wrapping it in a thin agent harness with real tools so the comparison is system vs system not model vs model"  
[X Link](https://x.com/anayatkhan09/status/2002438753064005725)  2025-12-20T17:59Z [---] followers, [---] engagements


"Week [--] of learning the Vercel AI SDK This week was about moving from chat responses real agent execution 🧵"  
[X Link](https://x.com/anayatkhan09/status/2002680479288369377)  2025-12-21T10:00Z [---] followers, [---] engagements


"1. Streaming is more than text - Streams include steps tool inputs tool outputs and text - The UI reflects what the agent is doing not just what it says"  
[X Link](https://x.com/anayatkhan09/status/2002680480844390468)  2025-12-21T10:00Z [---] followers, [--] engagements


"2. Tool calling = real actions - Agents can read write and update files - Each tool call becomes its own execution step"  
[X Link](https://x.com/anayatkhan09/status/2002680483188973948)  2025-12-21T10:00Z [---] followers, [--] engagements


"3. Making tools visible in the UI - Tool results can be rendered directly - Files written paths read outputs shown The UI becomes an agent debugger"  
[X Link](https://x.com/anayatkhan09/status/2002680484149485938)  2025-12-21T10:00Z [---] followers, [--] engagements


"Day [--] of learning the Vercel AI SDK Today I explored calling MCP servers via stdio. Key points: - MCP exposes external toolsets as agent tools - Agents can call real systems like GitHub via MCP servers - Stdio transport runs the MCP server as a local process - Tools are fetched dynamically and passed into streamText - Client must be closed explicitly after execution"  
[X Link](https://x.com/anayatkhan09/status/2003163849852354698)  2025-12-22T18:00Z [---] followers, [---] engagements


"@irl_danB The call stack framing for long running agents is spot on the missing piece is an execution trace object per subtask so you can compact by task boundary instead of truncating a flat chat log every N tokens"  
[X Link](https://x.com/anayatkhan09/status/2003353087583600799)  2025-12-23T06:32Z [---] followers, [--] engagements


"@VittoStack The only way this works at scale is stop letting AI write everything and start constraining it to boring scaffolding and glue code while humans own invariants error paths and the bits that wake you up at 3am"  
[X Link](https://x.com/anayatkhan09/status/2003355900208369880)  2025-12-23T06:44Z [---] followers, [--] engagements


"@maxedapps AI as assistants not vibe coding is the key line here Once youre wiring agents into real systems you see the hard parts are evals observability and failure modes not just generate code thats where dev skills compound with AI instead of competing with it"  
[X Link](https://x.com/anayatkhan09/status/2003511416536596756)  2025-12-23T17:01Z [---] followers, [---] engagements


"@github @AnthropicAI @code Agent Skills as an open standard is a huge win the missing piece now is a local skills registry so the same skill file runs in Copilot Claude Vercel AI SDK and any TypeScript agent runtime with zero rewrite"  
[X Link](https://x.com/anayatkhan09/status/2003512610944352563)  2025-12-23T17:06Z [---] followers, [---] engagements


"@omarsar0 The reality is this report shows most teams are gluing agent frameworks together anyway so Im leaning toward a thin TypeScript layer over raw APIs plus LangGraph style orchestration instead of betting on any single end to end framework"  
[X Link](https://x.com/anayatkhan09/status/2003514283687313452)  2025-12-23T17:13Z [---] followers, [--] engagements


"@corbin_braun AI vibe coding works here because you already knew how to wire [--] tier subscriptions plus per model pricing plus [--] webhooks the real win is using models to generate the boring idempotency and migration scripts while you sanity check the flows"  
[X Link](https://x.com/anayatkhan09/status/2003731019317543426)  2025-12-24T07:34Z [---] followers, [---] engagements


"@levie Two hours for two test cases is why we ended up encoding context policy as config per customer: explicit knobs for latency budget max hops and risk tier so the agent can choose retrieval vs internal on the fly"  
[X Link](https://x.com/anayatkhan09/status/2003731248615948468)  2025-12-24T07:35Z [---] followers, [---] engagements


"Day [--] of learning the Vercel AI SDK Context engineering Key points: - Prompt order matters. - Start with role + task context - Put background data in the middle - End with the actual ask rules and output format"  
[X Link](https://x.com/anayatkhan09/status/2003865563471524341)  2025-12-24T16:29Z [---] followers, [--] engagements


"@skirano [---] tool call loops is wild curious how youre handling failure modes and partial progress so the agent doesnt silently burn through the whole loop budget on bad branches"  
[X Link](https://x.com/anayatkhan09/status/2003867040315257056)  2025-12-24T16:35Z [---] followers, [----] engagements


"@backpinelabs The UX Agent and Product Engineer split is nice because you can enforce that every proposed view has a concrete /api/query behind it otherwise agents will happily invent joins that fall apart once you hit real prod cardinalities"  
[X Link](https://x.com/anayatkhan09/status/2004076152403681402)  2025-12-25T06:26Z [---] followers, [---] engagements


"@omarsar0 Using Claude Code the same way here but the real unlock was treating it as the agent harness and piping outputs through a thin TypeScript layer so tools file ops and retries stay consistent across models"  
[X Link](https://x.com/anayatkhan09/status/2004076263838015790)  2025-12-25T06:26Z [---] followers, [---] engagements


"@koylanai @replicate Context engineering as the separator between demo agents and production agents matches what we see too most failures are not the model but missing precedent around auth rate limits and partial failures encoded into skills"  
[X Link](https://x.com/anayatkhan09/status/2004076398601031850)  2025-12-25T06:27Z [---] followers, [---] engagements


"@rryssf_ Love the Task-Level Scheduled Sampling idea curious how robust Reason2Decide stays when you deploy on a shifted distribution of nurse notes and the models own prediction errors no longer match the training-time error profile"  
[X Link](https://x.com/anayatkhan09/status/2004077340771733991)  2025-12-25T06:30Z [---] followers, [---] engagements


"Day [--] of learning the @aisdk Today I wrote a full system prompt for a web agent. Key points: - good agents arent smart by default theyre constrained by design - clear capabilities cost-aware tools explicit permissions and output rules matter more than clever wording"  
[X Link](https://x.com/anayatkhan09/status/2004208131040333974)  2025-12-25T15:10Z [---] followers, [---] engagements


"Prompts https://github.com/agentcn-kit/ai-sdk-lab/blob/main/context-engineering/web/README.md https://github.com/agentcn-kit/ai-sdk-lab/blob/main/context-engineering/web/README.md"  
[X Link](https://x.com/anayatkhan09/status/2004208133993123926)  2025-12-25T15:10Z [---] followers, [--] engagements


"@CloudTrader4 @MiniMax__AI @velo_xyz Love how you wired M2.1 to Velo via get_market_context curious if you have a hard cap on tool frequency so the agent does not spam live exchange calls when users start iterating on the same query"  
[X Link](https://x.com/anayatkhan09/status/2004214994930749670)  2025-12-25T15:37Z [---] followers, [--] engagements


"@liambai21 Mapping atoms to bits is the interesting constraint here feels like the real leverage will come from agentic lab stacks where LLMs orchestrate liquid handlers and instruments over a proper lab state model not just read PDFs"  
[X Link](https://x.com/anayatkhan09/status/2004215867912163503)  2025-12-25T15:41Z [---] followers, [---] engagements


"@scaling01 Vibe coding gets really fun when Antigravitys plan file and comments become the main interface and the CLI is just a thin execution layer otherwise you lose all the shared context once the terminal scrolls"  
[X Link](https://x.com/anayatkhan09/status/2004470941070626831)  2025-12-26T08:34Z [---] followers, [---] engagements


"@danlovesproofs Benchmarks are useless only because they ignore real user journeys the fastest way to make this legible is pick one concrete workflow per category and record where each AI product actually fails in context not in a sandbox"  
[X Link](https://x.com/anayatkhan09/status/2004471284240175116)  2025-12-26T08:36Z [---] followers, [---] engagements


"@ryanvogel The wild part is your opencode vs cursor bill probably has less to do with Opus [---] and more with how aggressive the client is on tools context window and thinking flags a small change in default system prompt can double token burn on the same task"  
[X Link](https://x.com/anayatkhan09/status/2004472071905218859)  2025-12-26T08:39Z [---] followers, [---] engagements


"@pvncher The 200k tokens point is exactly why good AI setups push context into long lived artifacts like RFCs and decision logs the model just stitches code around those constraints instead of pretending to hold the whole product in its head"  
[X Link](https://x.com/anayatkhan09/status/2004472840884744651)  2025-12-26T08:42Z [---] followers, [---] engagements


"@0xPaulius Love this the next step for a claude code skills marketplace UI is a simple rating and telemetry layer so agents can auto promote skills that actually close tickets instead of whatever got installed the most today"  
[X Link](https://x.com/anayatkhan09/status/2004474154301681855)  2025-12-26T08:47Z [---] followers, [--] engagements


"@irl_danB The interesting thing about maximizing cache utilization is that in long running agents the real unit of cost is corrective intelligence not tokens so every stale cached summary that slips in quietly adds future review and refactor passes to the bill"  
[X Link](https://x.com/anayatkhan09/status/2004474340138709406)  2025-12-26T08:48Z [---] followers, [--] engagements


"@DataChaz @ycombinator YC vibe coding guide is great but the real unlock is turning those principles into agent readable skills so Claude or Cursor can actually execute version control planning and review loops instead of hoping the human remembers the checklist mid flow"  
[X Link](https://x.com/anayatkhan09/status/2004475317692645462)  2025-12-26T08:52Z [---] followers, [---] engagements


"@chatgpt21 The funny thing is that when you actually wire ChatGPT into tools and data the behavior looks way less like a database lookup and way more like a messy junior dev predicting the next token while juggling logs specs and partial context"  
[X Link](https://x.com/anayatkhan09/status/2004476132436103446)  2025-12-26T08:55Z [---] followers, [--] engagements


"@paulabartabajo_ Fine tuning a small model for browser control with GRPO is exactly the kind of thing that makes UI agents feel real curious how you are encoding reward beyond success fail so it learns to avoid flaky flows and dark patterns over time"  
[X Link](https://x.com/anayatkhan09/status/2004476911083548672)  2025-12-26T08:58Z [---] followers, [--] engagements


"@VitalikButerin @grok The interesting part about calling Grok inline on X is exactly that unpredictability once you cant script the answer ahead of time it behaves more like an external judge wired to real data than another quote tweet for your side"  
[X Link](https://x.com/anayatkhan09/status/2004477485917081817)  2025-12-26T09:00Z [---] followers, [--] engagements


"@pmddomingos RL only really shines over this kind of reward to every action supervised setup when exploration or off policy correction actually matter most of the agent stacks that ship today are closer to noisy behavior cloning plus value heads than classic RL anyway"  
[X Link](https://x.com/anayatkhan09/status/2004477652800032788)  2025-12-26T09:01Z [---] followers, [---] engagements


"@mathew_fuller @koylanai teach agents a simple pattern there: store decision traces as stateful policies then have a meta policy that scores each new situation on similarity vs novelty and only lets the agent break precedent when its uncertainty or expected upside crosses a set threshold"  
[X Link](https://x.com/anayatkhan09/status/2004480234545164430)  2025-12-26T09:11Z [---] followers, [--] engagements


"@phototypelabs For anyone building on multiple providers the real test is how fast Gemini [--] shows up in SDKs and infra where teams already live not just Studio UX"  
[X Link](https://x.com/anayatkhan09/status/2004563459384766852)  2025-12-26T14:42Z [---] followers, [---] engagements


"@techNmak Most stacks treating memory as a pile of embeddings is also why agents hallucinate state; the hard part is syncing a semantic layer like Cognee with real source of truth systems without creating a second drifting brain"  
[X Link](https://x.com/anayatkhan09/status/2004565054700818908)  2025-12-26T14:48Z [---] followers, [---] engagements


"Day [--] of learning the @aisdk Today I explored Evalite basics Evals = data + task + scorer. Instead of pass or fail you get a [----] score that shows how well your system behaves. This is how LLM behavior becomes testable repeatable and debuggable"  
[X Link](https://x.com/anayatkhan09/status/2004572246459703541)  2025-12-26T15:17Z [---] followers, [---] engagements


"@ManusAI @github Curious how you handle conflicts in the two way sync when Manus regenerated code touches the same regions as local commits; are you doing file level or AST level merges before pushing back to GitHub"  
[X Link](https://x.com/anayatkhan09/status/2004589043649278068)  2025-12-26T16:24Z [---] followers, [---] engagements


"@steipete codexbar feels like the perfect case study for why model level lock in hurts dev tooling would be cool if there was a thin provider interface so people could swap codex [---] for local or OSS models without touching the menu bar UX"  
[X Link](https://x.com/anayatkhan09/status/2004589663420567697)  2025-12-26T16:26Z [---] followers, [--] engagements


"@iannuttall Hackers shouldnt be able to get back in is exactly why Im nervous relying on gpt [---] as the security boundary feels safer to treat the model as an assistant to a proper policy engine and infra level controls"  
[X Link](https://x.com/anayatkhan09/status/2004589958636753120)  2025-12-26T16:27Z [---] followers, [---] engagements


"@thsottiaux Interfaces are lagging the capability jump for sure especially once you have agents orchestrating tools and long running flows; the hard bit now is designing interaction loops that expose planning state without overwhelming the user"  
[X Link](https://x.com/anayatkhan09/status/2004824379075826099)  2025-12-27T07:59Z [---] followers, [---] engagements


"@JacksonKernion Shaping LLM personality feels like the most tractable frontier from the trenches right now especially once you start wiring long lived agents where consistency over weeks matters more than raw capability jumps"  
[X Link](https://x.com/anayatkhan09/status/2004825167235875092)  2025-12-27T08:02Z [---] followers, [---] engagements


"@addyosmani The part about boring foundations turning agents from chaos generators into productivity multipliers is spot on; every time I skip tests or CI with AI generated changes the cleanup tax later easily erases the initial 10x feeling"  
[X Link](https://x.com/anayatkhan09/status/2004825969924997207)  2025-12-27T08:05Z [---] followers, [--] engagements


"@banteg Love the council of [--] agents idea the annoying part in practice is handling when amp gemini claude disagree on a chunk and you need a tie-breaker policy that does not stall a [--] hour long codex run"  
[X Link](https://x.com/anayatkhan09/status/2004826102133670004)  2025-12-27T08:06Z [---] followers, [--] engagements


"@mcuban The tech to create agents is easy the hard part in SMBs is gating those small mistakes at scale behind human review queues so you do not silently ship a thousand slightly wrong invoices overnight"  
[X Link](https://x.com/anayatkhan09/status/2004826347152237039)  2025-12-27T08:07Z [---] followers, [--] engagements


"@RyanShowlive @mcuban Agents flip that math only if you plug into billing and bank feeds as source of truth and treat every invoice decision as a logged event otherwise you just move those small mistakes at scale into a black box"  
[X Link](https://x.com/anayatkhan09/status/2004826414751826078)  2025-12-27T08:07Z [---] followers, [---] engagements


"@vasuman The real gap between GPT 4o in Windsurf and Opus [---] in Claude Code is usually the harness and workflow not the weights; once you add planning file context and strict edit diffs most models suddenly feel way more senior"  
[X Link](https://x.com/anayatkhan09/status/2004826631316402499)  2025-12-27T08:08Z [---] followers, [---] engagements


"@corbin_braun AI vibe coding only works when you keep a tight mental model of the system shape otherwise rate limiting and webhook flows like your Stripe email queue turn into silent failure factories a week after shipping"  
[X Link](https://x.com/anayatkhan09/status/2004828620460884130)  2025-12-27T08:16Z [---] followers, [---] engagements


"Day [--] of learning the @aisdk Today I explored deterministic evals. Instead of using another LLM to judge outputs you score them in code. Check things like: markdown links included output length limits Fast predictable unit-teststyle evals for LLM behaviour"  
[X Link](https://x.com/anayatkhan09/status/2004921158504743077)  2025-12-27T14:23Z [---] followers, [---] engagements


"@0xSero You want the AI output to match your brand is the key. The real unlock is wiring that branding prompt into your actual stack so every agent and UI generator calls the same style system instead of pasting prompts per project"  
[X Link](https://x.com/anayatkhan09/status/2004930347335815210)  2025-12-27T15:00Z [---] followers, [---] engagements


"@AlemTuzlak @tan_stack The most powerful API ever framing is relatable. The hard part I keep hitting with do everything AI APIs is not power but debuggability and testability you need boring primitives if you want teams to ship on top of it"  
[X Link](https://x.com/anayatkhan09/status/2004933449766109476)  2025-12-27T15:12Z [---] followers, [--] engagements


"@thsottiaux Background Terminal sounds perfect for long running test suites. Curious if you are planning any resource budgeting so the model does not over fan out shells and choke on I O or CPU on smaller dev machines"  
[X Link](https://x.com/anayatkhan09/status/2005162423838449741)  2025-12-28T06:22Z [---] followers, [----] engagements


"@FinnMurphy12 The funny part is we already see the limits when you try to bolt RL onto todays models for real products. Getting stable reward signals from messy human UX is 100x harder than spinning up another frontier checkpoint"  
[X Link](https://x.com/anayatkhan09/status/2005162577521951113)  2025-12-28T06:23Z [---] followers, [----] engagements


"@Simeon_Cps Terminal-only Claude feels fine until an agent silently refactors a cross-cutting concern and your mental model lags behind. Tight loop is Claude drives changes IDE/git diff drives your attention otherwise you only notice coupling bugs at PR time"  
[X Link](https://x.com/anayatkhan09/status/2005162690667503699)  2025-12-28T06:23Z [---] followers, [--] engagements


"@Angaisb_ Putting current time in the system prompt is a KV cache killer but you can still get most of the UX win by injecting a normalized time block into the first user message of a session and letting tools handle precision when needed"  
[X Link](https://x.com/anayatkhan09/status/2005163602198888576)  2025-12-28T06:27Z [---] followers, [--] engagements


"@corbin_braun The model that can code anything is only an edge if you wire it into a tight loop with tests tracing and rollback. Raw Opus [---] is insane but the moat is the workflow you wrap around it not the API key"  
[X Link](https://x.com/anayatkhan09/status/2005164153812713595)  2025-12-28T06:29Z [---] followers, [---] engagements


"@petergyang @claudeai Forking a conversation is great UX but the hard part is how much context to clone. The sweet spot feels like forking from a message plus a small summary window then letting retrieval bring in older chunks on demand"  
[X Link](https://x.com/anayatkhan09/status/2005164514954936731)  2025-12-28T06:30Z [---] followers, [---] engagements


"@JasonBotterill Inference speeds getting so quick that reasoning feels instant only matters if your tools and storage keep up. Right now most of the latency is in janky toolchains slow DBs and flaky browsers not the forward pass"  
[X Link](https://x.com/anayatkhan09/status/2005165723652657551)  2025-12-28T06:35Z [---] followers, [---] engagements


"@aisdk [--]. MCP as external toolsets Agents can call real systems (GitHub) via MCP Tools dont have to live inside your codebase"  
[X Link](https://x.com/anayatkhan09/status/2005285165967310902)  2025-12-28T14:30Z [---] followers, [--] engagements


"@aisdk [--]. Chat persistence is built in useChat sends a UUID by default You can override it with a chatId from the URL This enables durable chats reload-safe state and real conversation continuity"  
[X Link](https://x.com/anayatkhan09/status/2005285167410069934)  2025-12-28T14:30Z [---] followers, [--] engagements


"@aisdk [--]. Context engineering beats clever prompts Prompt order matters Start with role + task context Put background data in the middle End with rules the ask and output format"  
[X Link](https://x.com/anayatkhan09/status/2005285169851154718)  2025-12-28T14:30Z [---] followers, [--] engagements


"@johnlindquist Agent smell hits hardest when it hallucinates Vercel AI SDK streaming patterns that break on edge cases like aborted requests. Fixed by pinning tool calls in the loop but now prompts balloon.x"  
[X Link](https://x.com/anayatkhan09/status/2005296906709487738)  2025-12-28T15:16Z [---] followers, [--] engagements


"@itsmechase Composer-1 handles boilerplate fine until TypeScript generics get nested three levels deep and it drops coherence. Switched to Vercel AI SDK streaming for agent loops where model tiering actually sticks"  
[X Link](https://x.com/anayatkhan09/status/2005297235467399273)  2025-12-28T15:18Z [---] followers, [----] engagements


"@donvito Sandbox escapes state loss when Claude restarts mid-generation on Vercel AI SDK agent handoffs. Local restrictions help but kill persistence across sessions"  
[X Link](https://x.com/anayatkhan09/status/2005298488171810865)  2025-12-28T15:23Z [---] followers, [---] engagements


"@gfodor Ownership stake explains it but even then AI coding kills the feedback loop on edge cases like Vercel AI SDK stream handling where models hallucinate state bugs you only catch debugging live traffic"  
[X Link](https://x.com/anayatkhan09/status/2005546807317938332)  2025-12-29T07:49Z [---] followers, [---] engagements


"@SCHIZO_FREQ Claude code shines on greenfield but 80% job debunk assumes no infra debt; refactoring agent state machines in Vercel AI SDK often needs [--] human passes to stop silent race conditions under load"  
[X Link](https://x.com/anayatkhan09/status/2005546977979965739)  2025-12-29T07:50Z [---] followers, [---] engagements


"@RayFernando1337 Hooks in Cursor are great until agent loops hit a type-regen failure and cascade frontend breaks. Switched to Vercel AI SDK tool calls with inline validation to catch sync issues mid-run instead of post-file-change. Less brittle for TypeScript stacks"  
[X Link](https://x.com/anayatkhan09/status/2005547139972391248)  2025-12-29T07:51Z [---] followers, [---] engagements


"@PaulSolt Codex accelerates prototypes but hallucinates state management in agent loops until you wire explicit checkpoints. Seeing that tradeoff daily in Vercel AI SDK builds"  
[X Link](https://x.com/anayatkhan09/status/2005547482621886949)  2025-12-29T07:52Z [---] followers, [--] engagements


"@omarsar0 Single ReAct internalizing complexity hits perf but explodes on non-stationary domains like fast-moving repos or unpublished infra. Atomic actions drift without external memory hooks into vector stores or live APIs"  
[X Link](https://x.com/anayatkhan09/status/2005547830862373362)  2025-12-29T07:53Z [---] followers, [--] engagements


"@steipete Single context works until task queue hits 128k where model starts hallucinating non-existent pain points from token overflow. Switched to Vercel AI SDK streaming to chunk refactor suggestions live"  
[X Link](https://x.com/anayatkhan09/status/2005548159720927555)  2025-12-29T07:55Z [---] followers, [---] engagements


"@BenjaminDEKR Claude Code scales ambition until integration loops where generated deps conflict with your TypeScript monorepo constraints. Real OS needs human-orchestrated agent swarms to resolve live"  
[X Link](https://x.com/anayatkhan09/status/2005548268529619237)  2025-12-29T07:55Z [---] followers, [---] engagements


"@EricTopol Infra is primed for generative AI takeoff. Real question: will users sustain adoption like social media GenZ/Alpha will cheap access plus craving virtual escapes over real-world reality"  
[X Link](https://x.com/anayatkhan09/status/2005550017764049078)  2025-12-29T08:02Z [---] followers, [--] engagements


"@CloudTrader4 Terminal UI skills shine for agent scaffolding but token bloat kills perf when chaining LLM calls through TUI state machines. Need streaming diffs to keep loops live"  
[X Link](https://x.com/anayatkhan09/status/2005550313542213792)  2025-12-29T08:03Z [---] followers, [---] engagements


"@lucas_montano AI writes code but hallucinates Vercel AI SDK hooks that break TypeScript monorepo builds at deploy. Understanding diffs catches 80% of prod failures before they ship"  
[X Link](https://x.com/anayatkhan09/status/2005550568476180956)  2025-12-29T08:04Z [---] followers, [--] engagements


"@petergyang @cursor_ai Cursor md preview exists but context switching to Obsidian kills agentic workflows when chaining Vercel AI SDK docs into live TypeScript edits. Native streaming md would fix token loss on reloads"  
[X Link](https://x.com/anayatkhan09/status/2005550844012630430)  2025-12-29T08:05Z [---] followers, [---] engagements


"@damianplayer Vertical agents fail when industry APIs change without notice breaking 80% code reuse assumption. Need Vercel AI SDK style tool calling with live schema validation to survive prod drift"  
[X Link](https://x.com/anayatkhan09/status/2005551541646958617)  2025-12-29T08:08Z [---] followers, [--] engagements


"@gregisenberg Tried Claude Code for React Native MVP last week. State management explodes past day [--] without manual hooks even with edge case prompts"  
[X Link](https://x.com/anayatkhan09/status/2005551992694001767)  2025-12-29T08:10Z [---] followers, [---] engagements


"@NoahEpstein_ n8n plain english workflows break when external APIs change schemas mid-month. Retries eat credits fast without baked-in versioning"  
[X Link](https://x.com/anayatkhan09/status/2005552565413675455)  2025-12-29T08:12Z [---] followers, [--] engagements


"@brunoborges LLMs amplify incoherent abstractions exactly like you said. Agent handoffs in my TypeScript flows turn into 20% perf regressions without strict schema contracts upfront"  
[X Link](https://x.com/anayatkhan09/status/2005552784327024928)  2025-12-29T08:13Z [---] followers, [---] engagements


"@gdb Coding agents ship fast until schema drift hits in multi-step TypeScript chains. Fixed yesterday by pinning Vercel AI SDK versions across handoffs"  
[X Link](https://x.com/anayatkhan09/status/2005553457890234784)  2025-12-29T08:16Z [---] followers, [---] engagements


"Day [--] of learning the @aisdk Today I explored LLM-as-a-judge evals. Instead of hard rules in code another model evaluates the output. You define: a scoring schema structured feedback grade numeric score mapping"  
[X Link](https://x.com/anayatkhan09/status/2005665173940400597)  2025-12-29T15:40Z [---] followers, [----] engagements


"@jerryjliu0 Filesystem tools are great until latency and search UX fall apart once youre above small-to-medium. The interesting frontier is hybrid: agent uses the FS for traversal but maintains cheap task-specific indexes for hot paths"  
[X Link](https://x.com/anayatkhan09/status/2005894189612216461)  2025-12-30T06:50Z [---] followers, [---] engagements


"@jasonzhou1993 Same itch here using Cursor mostly as a diff viewer for Claude; the missing piece is a terminal first layout where the agent can drive git tests and editors as first class panes not plugins"  
[X Link](https://x.com/anayatkhan09/status/2005896138315161812)  2025-12-30T06:58Z [---] followers, [---] engagements


"@thsottiaux Once laptops are no longer the bottleneck the constraint is git and terminals that serialize everything; been hacking on agent native flows where merges tests and deploys all run in parallel sandboxes instead of a single shell"  
[X Link](https://x.com/anayatkhan09/status/2005896612837544235)  2025-12-30T06:59Z [---] followers, [---] engagements


"@rauchg The wild part in practice is that once agents write and run the tests most regressions shift from missing coverage to bad specs; the hard work becomes tightening contracts and fixtures not cranking out more green checks"  
[X Link](https://x.com/anayatkhan09/status/2005904460636139926)  2025-12-30T07:31Z [---] followers, [--] engagements


"@jefftangx The export and then I can no longer iterate inside Manus bit is the real moat gap. Until the sandbox has first class git env secrets and deploy targets complex apps will keep leaking out into local dev and other agents"  
[X Link](https://x.com/anayatkhan09/status/2006036028449079788)  2025-12-30T16:13Z [---] followers, [--] engagements


"@t_blom Parallel phone enquiries with voice AI only DDOS if businesses stay human-first. The interesting design problem is rate limiting and per-caller identity so agents can fan out without turning into robocall spam"  
[X Link](https://x.com/anayatkhan09/status/2006036215745683820)  2025-12-30T16:14Z [---] followers, [---] engagements


"@dani_avila7 @hedgineering The multi domain chaining is super interesting especially UI Skills calling into Data and Infra. Curious how you avoid Skills fighting each other on bigger refactors where multiple domains all think they own the change"  
[X Link](https://x.com/anayatkhan09/status/2006269789908676901)  2025-12-31T07:42Z [---] followers, [---] engagements


"@levie The tech being just hard enough is true but the real moat is boring stuff like RBAC audit trails and incident playbooks for agents touching revenue systems. That glue is what turns a demo into an enterprise product"  
[X Link](https://x.com/anayatkhan09/status/2006269940765212721)  2025-12-31T07:43Z [---] followers, [---] engagements


"@arnaudai The daily regeneration part is fun but the real unlock is wiring that markdown spec into analytics and state so layout and CTAs adapt to who is visiting and what they did last session not just a fresh coat of gradients"  
[X Link](https://x.com/anayatkhan09/status/2006271983731016187)  2025-12-31T07:51Z [---] followers, [---] engagements


"@burkov Pattern following is exactly why agents are scary good at the boring [--] percent and terrible at the last 20%. The real job now is picking constraints and feedback loops so the pattern machine does not quietly cement bad designs"  
[X Link](https://x.com/anayatkhan09/status/2006272593637302769)  2025-12-31T07:53Z [---] followers, [---] engagements


"@harveyhucal Login is the easy part the scary part is long lived auth plus agents that can schedule actions across many tenants. Curious how you sandbox workflows so one prompt cannot accidentally blast the wrong account or environment"  
[X Link](https://x.com/anayatkhan09/status/2006272865197514925)  2025-12-31T07:55Z [---] followers, [--] engagements


"@GoldenAIdev @aisdk Start with Evalites to get the basics right then use @langfuse for executing and tracking evals. https://www.evalite.dev/ https://www.evalite.dev/"  
[X Link](https://x.com/anayatkhan09/status/2006411790385913875)  2025-12-31T17:07Z [---] followers, [--] engagements


"@gdb Enterprise agent adoption happens when the agent can safely touch prod systems not just dashboards and docs. The hard part is wiring identity permissions and observability into every tool call so ops actually trust it"  
[X Link](https://x.com/anayatkhan09/status/2006623200478900729)  2026-01-01T07:07Z [---] followers, [---] engagements


"@bentossell AI IDE vs AI CLI feels like a false binary. The setups that scale are CLI driven agents with enough workspace awareness and UI affordances to debug multi file refactors without losing the speed of typing one command"  
[X Link](https://x.com/anayatkhan09/status/2006623872733557204)  2026-01-01T07:09Z [---] followers, [--] engagements


"@jukan05 A [----] dollar per month AI agent only makes sense if it is wired directly into revenue levers. The real question for [----] is who ships the infra that lets those agents safely own P L without a human in the loop on every decision"  
[X Link](https://x.com/anayatkhan09/status/2006624090262852004)  2026-01-01T07:10Z [---] followers, [--] engagements


"@connordavis_ai The funny part is most real systems already do Meaning Task Process just implicitly in glue code. The win is when that semantic layer is a first class module you can unit test and reuse across agents instead of redoing prompt gymnastics every time"  
[X Link](https://x.com/anayatkhan09/status/2006624214024073356)  2026-01-01T07:11Z [---] followers, [---] engagements


"@dashboardlim Most of this playbook works until the client wants one-off exceptions or new channels. The agencies that stick around are the ones who turn that signature workflow into a parameterized product with proper logging and change management"  
[X Link](https://x.com/anayatkhan09/status/2006624907669680528)  2026-01-01T07:13Z [---] followers, [--] engagements


"@ekuyda @Replit @amjad Consumer vibe coding is harder mostly because there is no concept of workspace or state yet. The moment tools expose project memory history and safe actions to non devs you get real consumer platforms not just one off magic prompts"  
[X Link](https://x.com/anayatkhan09/status/2006625852977721555)  2026-01-01T07:17Z [---] followers, [--] engagements


"@MervinPraison The real unlock here is pairing GLM [---] with a watchdog agent. Long running coding sessions only feel safe when you checkpoint filesystem state cap diff size per step and auto roll back if tests or lints start failing"  
[X Link](https://x.com/anayatkhan09/status/2006626733102150131)  2026-01-01T07:21Z [---] followers, [---] engagements


"@GergelyOrosz The cool part in your backlog story is that agents are finally good enough to own the boring scaffolding. The next step is letting them also spin up infra and CI stubs by default so side projects are deployable not just running on localhost"  
[X Link](https://x.com/anayatkhan09/status/2006627178713407570)  2026-01-01T07:22Z [---] followers, [--] engagements


"@joshpuckett Vibe coded apps like this are the best demo that front end is still where models struggle most. The fun part is teaching the agent your motion design system once so every new app inherits the same feel by default"  
[X Link](https://x.com/anayatkhan09/status/2006627772261896674)  2026-01-01T07:25Z [---] followers, [--] engagements


"Im tired of rebuilding auth + chat UI for every agent. So I'm building an open-source AgentKit starter with: - landing page ( @magicuidesign @framer ) - auth + DB ( @clerk @supabase @prisma ) - chat UI ( @Ibelick's lib ) - @aisdk integration - analytics ( @posthog ) - real agent patterns not toy examples Releasing this week. https://twitter.com/i/web/status/2006785835405029649 https://twitter.com/i/web/status/2006785835405029649"  
[X Link](https://x.com/anayatkhan09/status/2006785835405029649)  2026-01-01T17:53Z [---] followers, [---] engagements


"@ivanburazin Sandboxes as the primitive makes sense but the sharp edge is observability. Once agents live in opaque sandboxes you need first-class traces snapshots and replay or you just moved the chaos one layer down"  
[X Link](https://x.com/anayatkhan09/status/2007061928515117153)  2026-01-02T12:10Z [---] followers, [--] engagements


"@AlexFinn The hard part is definitely not done. The actual multi trillion opportunity is giving Opus boring capabilities like durable memory resource limits and safe tool routing so it can run unattended without nuking a codebase or a PnL"  
[X Link](https://x.com/anayatkhan09/status/2007062198544445616)  2026-01-02T12:11Z [---] followers, [--] engagements


"@learn2vibe This learn from your preferences pattern is strong the missing piece is persistence. Wiring the chosen variants back into a tiny preference schema or skill file makes Claudes taste actually compound across projects instead of resetting every session"  
[X Link](https://x.com/anayatkhan09/status/2007064196710895823)  2026-01-02T12:19Z [---] followers, [---] engagements


"@raroque Love this web ui on top of a local Claude Code subscription is the sweet spot. Gives you all the agent IDE ergonomics without handing repo and system access to some random hosted black box"  
[X Link](https://x.com/anayatkhan09/status/2007068594421813280)  2026-01-02T12:36Z [---] followers, [---] engagements


"@mert The part about doing anything with some high level system design guidance is the trap; the hard part is turning that into robust agent workflows with idempotent tools retries and telemetry so it doesnt silently ship broken systems"  
[X Link](https://x.com/anayatkhan09/status/2007344194277572927)  2026-01-03T06:52Z [---] followers, [--] engagements


"@petergyang The focus on enterprise and coding is real but the secret sauce is how deep they went on agent affordances in Claude Code and MCP instead of just shipping another chat UI with a code interpreter badge"  
[X Link](https://x.com/anayatkhan09/status/2007344673854276012)  2026-01-03T06:53Z [---] followers, [---] engagements


"@jarrodwatts Fun until Opus [---] silently ships something slightly off spec; pairing it with a second model that only reviews diffs against a strict contract is the only way Ive kept long running repos from drifting"  
[X Link](https://x.com/anayatkhan09/status/2007344871984771233)  2026-01-03T06:54Z [---] followers, [--] engagements


"@Angaisb_ The funny part is that all the missing features and bad rate limits probably forced them to over invest in Claude Code MCP and agent ergonomics instead of shiny UX which is exactly what makes it feel so usable for real engineering work"  
[X Link](https://x.com/anayatkhan09/status/2007344963785552303)  2026-01-03T06:55Z [---] followers, [---] engagements


"@Yuchenj_UW Wild part for me is not Claude Code beating an internal tool it is that once you let engineers pick their own agentic stack you basically admit the constraint now is workflow design and evals not model ownership"  
[X Link](https://x.com/anayatkhan09/status/2007345322205581785)  2026-01-03T06:56Z [---] followers, [---] engagements


"@RhysSullivan Love that you want one repo for everything the hard part is making the life manager an actual event bus so the recipe and bookmarks agents can subscribe to the same state instead of each bolting on their own cron logic"  
[X Link](https://x.com/anayatkhan09/status/2007345441055354976)  2026-01-03T06:57Z [---] followers, [--] engagements


"@craigzLiszt Vibe coding is great for shipping fast but the edge is combining it with boring stuff like typed contracts tests and observability so you are not vibe debugging a prod incident at 3am"  
[X Link](https://x.com/anayatkhan09/status/2007346511219437816)  2026-01-03T07:01Z [---] followers, [--] engagements


"@RhysSullivan Running [--] Claude Code instances is fun until you realize none of them share state or constraints; the real flex is a coordinator that tracks tasks repos and budgets so agents are parallel but not fighting each other"  
[X Link](https://x.com/anayatkhan09/status/2007348708732723297)  2026-01-03T07:10Z [---] followers, [----] engagements


"@SherryYanJiang ASCII first is such a good constraint especially when you then have Opus generate a component tree and test ids from the terminal chart so the layout routing and E2E selectors all stay in sync as the UI evolves"  
[X Link](https://x.com/anayatkhan09/status/2007348786604191771)  2026-01-03T07:10Z [---] followers, [--] engagements


"@forgebitz The still pushing pixels part is why I treat CLI tools as layout engines only and then keep a tiny design system repo that Claude or Gemini can target so tweaks happen at the token level not per screen"  
[X Link](https://x.com/anayatkhan09/status/2007351115554431303)  2026-01-03T07:19Z [---] followers, [--] engagements


"@daniel_nguyenx The fun part in [----] is you can hand a lot of the boring scaffolding to Claude or Cursor but the real leverage to hit 10k MRR is wiring agents into onboarding support and expansion so your vertical SaaS basically runs growth experiments on autopilot"  
[X Link](https://x.com/anayatkhan09/status/2007352274847445253)  2026-01-03T07:24Z [---] followers, [--] engagements


"@doodlestein The remote Ubuntu plus iPhone screenshot flow is such a good niche makes me want to pair giil with a Claude Code skill that watches that folder and auto opens the right repo file and viewport to reproduce each glitch"  
[X Link](https://x.com/anayatkhan09/status/2007352433933201897)  2026-01-03T07:24Z [---] followers, [---] engagements


"Building an open-source AgentKit Todays progress: - Full chat UI - Theme switcher Laying down solid UX + infra before shipping real agents"  
[X Link](https://x.com/anayatkhan09/status/2007469818778116127)  2026-01-03T15:11Z [---] followers, [--] engagements


"@carlvellotti Claude being pretty darn slow is the real constraint here. The win is when you chain a few painful but low-context flows together in one shot refund plus FCC form plus cancel service not when you try to make it feel real time"  
[X Link](https://x.com/anayatkhan09/status/2007718517777584550)  2026-01-04T07:39Z [---] followers, [---] engagements


"@kr0der @steipete Codex yolo is fun until it silently edits outside the task surface. I always pair yolo with a sandbox branch plus a path allowlist so it literally cannot touch anything outside the current change set"  
[X Link](https://x.com/anayatkhan09/status/2007718922737578492)  2026-01-04T07:41Z [---] followers, [---] engagements


"@dani_avila7 Sandbox Claude Code cleaning like this is great but for real machines I would always start with a dry run listing every path it plans to touch so you can diff and veto before it rm -rfs your life"  
[X Link](https://x.com/anayatkhan09/status/2007719008620163282)  2026-01-04T07:41Z [---] followers, [---] engagements


"@rileybrown Running Claude Code as a general agent in a virtual terminal is nice but the real unlock is layering a tiny task router on top so long running ops do not block the same shell session"  
[X Link](https://x.com/anayatkhan09/status/2007719191667978681)  2026-01-04T07:42Z [---] followers, [---] engagements


"@deredleritt3r The primary purpose framing matters less in practice than the feedback loop. Once Claude Code and Codex started shipping everyday dev wins the volume of real world traces they get for AI research went vertical"  
[X Link](https://x.com/anayatkhan09/status/2007719823573418360)  2026-01-04T07:44Z [---] followers, [---] engagements


"@leerob The happy path is magic but once agents live in cloud sandboxes you really need guardrails like max token budgets and heartbeat checks so a stuck job does not quietly burn 50M tokens while you sleep"  
[X Link](https://x.com/anayatkhan09/status/2007720393608704077)  2026-01-04T07:46Z [---] followers, [--] engagements


"@rauchg @vercel @awscloud Routing Claude Code through AI Gateway is nice for centralizing spend"  
[X Link](https://x.com/anayatkhan09/status/2007720579613503852)  2026-01-04T07:47Z [---] followers, [--] engagements


"@craigzLiszt Focus helps but Claude Code is winning at generative software because the product is tuned around long lived coding workflows and repo scale context not just a better base model in isolation"  
[X Link](https://x.com/anayatkhan09/status/2007720725088784711)  2026-01-04T07:48Z [---] followers, [--] engagements


"@arvidkahl Claude Code on non trivial codebases works way better when you also bound the blast radius. I like your flow plus a final step where it only edits files already touched in the repro path"  
[X Link](https://x.com/anayatkhan09/status/2007720832714584436)  2026-01-04T07:48Z [---] followers, [--] engagements


"@skeptrune Cursor vs Claude Code is mostly about workflow shape for me. I reach for Claude when I need repo wide refactors and multi step plans Cursor when I want super tight inline edits in a single file"  
[X Link](https://x.com/anayatkhan09/status/2007720953569259976)  2026-01-04T07:49Z [---] followers, [--] engagements


"@GergelyOrosz Nondeterminism is only a blocker if the task has no cheap verifier. For most agentic systems that ship the real win is designing tight validators and idempotent tools so you can lean into best of N without trashing UX or data"  
[X Link](https://x.com/anayatkhan09/status/2007723535054320013)  2026-01-04T07:59Z [---] followers, [--] engagements


"@ahall_research Cool result but the interesting part for agents isnt that Claude Code wrote a paper its that it reliably orchestrated git data fetching stat tests and repo updates. Curious how often the tool calls silently failed on the way there"  
[X Link](https://x.com/anayatkhan09/status/2007723669561516273)  2026-01-04T07:59Z [---] followers, [---] engagements


"@theo Most of the token resellers people complain about are really UX resellers. The only ones that survive are the ones that own routing grounding memory and evals so swapping models becomes an implementation detail"  
[X Link](https://x.com/anayatkhan09/status/2007724262711537770)  2026-01-04T08:02Z [---] followers, [---] engagements


"@paulnovosad The half working prototypes part is so real. The only people getting durable wins with Claude Code are treating it like a junior pair programmer and enforcing boring stuff like tests simple architectures and explicit invariants between runs"  
[X Link](https://x.com/anayatkhan09/status/2007724699036565616)  2026-01-04T08:04Z [---] followers, [---] engagements


"@DeryaTR_ Agency is all you need until an agent has long term tool access and no safety rails. The boring work now is evals permissions and circuit breakers so game over moments do not become real world incidents"  
[X Link](https://x.com/anayatkhan09/status/2007726322303574382)  2026-01-04T08:10Z [---] followers, [--] engagements


"@hyperbrowser Context caching is nice but the real win is controlling granularity. The moment /docs fetch pulls a 1k line page into Claude you are fighting context bloat and noisy grounding unless you chunk diff and prune aggressively per repo"  
[X Link](https://x.com/anayatkhan09/status/2008160071759597726)  2026-01-05T12:54Z [---] followers, [---] engagements


"@Dobrenkz The HQ idea is right but the failure mode is one giant context blob. The setups that keep working are boring: per client repos hard file boundaries and agents that only ever see the slice they need"  
[X Link](https://x.com/anayatkhan09/status/2008160707259625501)  2026-01-05T12:56Z [---] followers, [---] engagements


"@boringmarketer Learning what broken looks like is the real unlock. The people who actually set up git env vars APIs and ship a tiny app with Claude Code end up miles ahead when they later use it for pure marketing work"  
[X Link](https://x.com/anayatkhan09/status/2008160793012183279)  2026-01-05T12:56Z [---] followers, [--] engagements


"@samgoodwin89 The part about needing everything on one page is exactly the tradeoff. GUI is unbeatable for reviewing diffs but the minute you want repeatable multi step agents a boring CLI pipeline with tests starts to win"  
[X Link](https://x.com/anayatkhan09/status/2008161718510518601)  2026-01-05T13:00Z [---] followers, [--] engagements


"@rileybrown The video game feeling is real but in teams the boring stuff still matters tests logs infra and rollback plans. Claude can speedrun a feature but someone still owns the wipe when prod goes sideways"  
[X Link](https://x.com/anayatkhan09/status/2008174612019048630)  2026-01-05T13:51Z [---] followers, [--] engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing