I’m starting to feel like AI agents are exposing a flaw that already existed in engineering teams: We review changes in isolation, but systems fail through interactions. Agents just scale that mismatch faster.

I'll be attending AI Engineer Singapore! If you're there too, let's catch up!

One thing I’ve noticed while working with coding agents: The issue usually isn’t “bad code.” It’s: > retries affecting another service > timeouts shifting load patterns > one “safe” local optimization causing global instability

Just asked Codex to add some features. After 10 mins:

We've been Building a new feature. Spent days overthinking it. - Workflows, - agents, - orchestration. Then we realized, just use tool calls! No agents needed. No extra layers. Just function calls. Works perfectly.

My @cursor_ai Agents SDK video is already ranking on Google. Working on a deeper breakdown now. Stay tuned.

Awesome AI Apps is Trending on GitHub! Thanks Everyone!

Shipping AI isn’t just writing prompts. You’re dealing with retries, latency, evals, cost, observability, and unpredictable failures. It’s still software engineering. Just less deterministic.

X is wild. One week everyone's hyping Claude Code. Next week they're all on Codex. I test them myself and stick with what actually works for my workflow. You should too. Don't follow the hype cycle.

Built a real-time voice agent that listens and responds instantly. - @SarvamAI STT for speech input - GPT-OSS via @nebiustf for processing - Sarvam TTS for voice output Wired everything together with @pipecat_ai into one streaming pipeline. Runs locally. No complex setup. Full tutorial:

MCP is useful. It makes integrations cleaner. But it doesn’t fix bad system design. You still need proper permissions, safe write paths, logging, and control over what the agent can do. It’s plumbing, not a safety layer.

YC is becoming the next JEE here!

The hard problem isn't storage. It's filtering. Engram by @weaviate_io solves it at ingestion: • Topics define what gets extracted • Scopes define where it lives Memory is structured before it hits the vector DB. No dumping raw logs and hoping retrieval fixes it later.

Quoting this because more people need to see how simple this gets. From blank UI → real-time collaborative editor with comments, presence, and live sync… without touching WebSockets or CRDTs. Here’s the demo 👇 x.com/Arindam_172...

Agents don’t become reliable because the model has improved. They become reliable when the system around them improves.

100% agree. We need very niche & specific benchmarks for each task/domain. That will help us to judge what works best for this use case. But my only concern is the fact that many teams will game the benchmarks. x.com/OfficialLog...

Had a great time speaking at @gdgkolkata's Build with AI event yesterday. Built a voice agent with web access using @livekit, @GeminiApp & @olostep. The energy in the room was amazing. People were genuinely excited about the possibilities. Here's the demo I showed:

I'll be speaking at Build with AI Kolkata by @gdgkolkata tomorrow. Will walk through building real-time voice agents with Gemini 3.1 Flash + Livekit that process audio directly for natural flow 📅 Apr 25 | 9:30 AM IST 📍 SNU, A Block Seminar Hall See you there!

New update for CC-lens 👀 Just shipped: • cmd + K navigation • cleaner session view • faster access to chats & tool calls UI feels way better now. Still 100% free & open source. try: npx cc-lens

I launched Formaly as a product for creating and running better surveys. Now I am thinking more about the next layer: what happens when this becomes part of a team workflow, not just a solo workflow? That shift changes the roadmap a lot.

When people say “multi-agent systems”, most examples are just prompt chains. A real multi-agent system needs: • orchestration • service boundaries • execution tracking • governance Otherwise, it becomes impossible to debug. Projects like @AgentField_ai are trying to formalize this stack.

Improved light mode for cc-lens. Send me feedback if you have any. Will keep improving it. It's 100% free & open source. Try: npx cc-lens

Just hit 2k subscribers on YouTube 🎉 Not a huge milestone, but it means a lot. Thanks for supporting the content. Gonna be way more consistent now.

Learn How to Build Real-Time Voice Agent With @GeminiApp 3.1 Flash Live & @livekit Watch it here: www.youtube.com/watch?v=edk... x.com/Arindam_172...

Function matters. Presentation matters too. I am improving how surveys look inside Formaly.io because better design changes how seriously people take the experience. Trust starts at the interface.

I keep thinking about this: feedback should happen where the user already is. Not every survey should live behind a separate link. I am exploring more in-product feedback workflows for Formaly because that feels like a much bigger opportunity.

Infra used to be built for developers. Now it’s slowly being rebuilt for agents. Agents are now writing and running software. They need - long-running compute, - sandboxes, and - fast model access. Agentic infrastructure will unlock a new class of AI-native products x.com/rauchg/stat...

Updated my GitHub README after a while. Thoughts? (shoutout @steipete for the inspo)

We instrument everything in production. Logs, traces, metrics, dashboards, profilers. Then we pay $50-200/month for an AI coding tool that generates detailed session logs... And never look at them. Now you can, Just Run npx cc-lens

Crossed 2k Weekly downloads on CC-lens! Great to see people actually using it! Try now > npx cc-lens

Everyone in my TL got into: → VibeCon by @emergentlabs → @ycombinator Startup School India And I didn't make either. Anyone willing to help me get into one of these? Would mean a lot.

Hit my rate limits twice on a single project. @bcherny @claudeai When's the excessive token usage getting fixed? I love Claude Code. Want to keep using it. But this makes it nearly impossible to ship. ( checked the usage graph on cc-lens )

Claude Code burns tokens like a black box. Built CC-lens so you can see where they actually go. Project spend, cache hits, tool usage, and full session replay. Recorded a full walkthrough if you want the details. x.com/Arindam_172...

Alibaba's Qwen 3.6-Plus just matched Claude Opus on coding. Half the size. 1M context. Built for agents. It's beating Kimi K2.5 and Gemini 3 Pro on SWE-Bench. Free on OpenRouter right now. Open source soon. Smaller models are finally catching up. x.com/Alibaba_Qwe...

Crossed 50 Github Stars! Thanks for showing love, everyone! I've pushed some improvements to it. Feel free to test it out! x.com/Arindam_172...

Claude Code bills you but tells you nothing. Introducing CC-lens. Your sessions. Your spending. Your cache efficiency, laid out in one dashboard. See exactly where your tokens go. Which projects burn budget. What actually caches. Just Run: npx cc-lens

Many developers confuse MCP servers and Agent Skills. Simple way to think about it: MCP → connects agents to external systems Agent Skills → teach agents how to perform tasks One exposes tools. The other encodes workflows. Together, they make agentic systems far more reliable.

Claude Code source just leaked on npm. Someone accidentally pushed a .map file to the public registry. The damage: → ~512K lines of code → ~1,900 files → Full application structure exposed This wasn't a hack. Just a bad deploy. HugOps to the Anthropic team. github.com/instructkr/...

Put your effort where it actually compounds.

Hey folks, Looking for video editors. Experience with tutorial-style or educational videos is a strong plus. If that sounds like you, send your previous work in DM.

One thing I care about with Formaly: not every survey should feel like a spreadsheet with a submit button. Sometimes a conversational flow gets better answers. Sometimes a classic form is the right call. That is why Formaly supports both.

A lot of building is invisible. - Not launches. - Not demos. Just edge cases, cleanup, naming, fixing flows, and making rough things feel obvious. That is a big part of what building Formaly actually looks like day to day.

One of my blogs ranked on page 1 of Google. Checked the stats: → 53,848 views from Google → 1,888 from Brave And more... Search traffic still works if you write what people actually search for.

G𝗼-𝘁𝗼 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗰𝗵𝗮𝗻𝗻𝗲𝗹𝘀 𝗳𝗼𝗿 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗮𝗻𝗱 𝘀𝘁𝗮𝘆𝗶𝗻𝗴 𝘀𝗵𝗮𝗿𝗽) 𝗶𝗻 𝗔𝗜

We just built a Startup Idea Validation Agent! Just drop your Startup idea & get: → Instant market insights → Competitive positioning → A full validation report Built with @tavily, @nebiustf, Google ADK & @streamlit. Demo here 👇

I think, It's time to take a break

Did you know this? Every Reddit thread is secretly a JSON API endpoint Here's how to print money with it:

I just deployed a model to a dedicated endpoint using @nebiustf The observability dashboard shows you everything you need: traffic volume, TTFT, error breakdowns, and prompt size distributions. All in real-time. Super simple to set up. Watch how it works:

I wrote a complete guide to LLM observability. What it is, why traditional APM isn't enough, the four pillars you need to instrument, and key metrics that matter. Covers simple deployments, RAG pipelines, and agentic systems. Link Below:

I just built a full-stack Reddit clone with Claude Code. Backend, database, auth, storage, deployment. Everything is done. One prompt. Fully autonomous. Here's how it worked, step by step: