Building AI agents in public. Sharing what breaks, what works, and the numbers behind both. Founder @Blevoapp
Joined April 2012
- Tweets1.1K
- Following653
- Followers1K
- Likes2.5K
Pinned Tweet
most AI agent demos die in week 2, not day 1.
5 reasons agents break in real use — from the builds I've watched:
🧵
what's the one task you stopped handing your agent because it was right most of the time but wrong when it counted?
opus 5.5 "laziness" isn't a regression. it's pricing.
the sailing test: gemini 4 argon won on a fifth of the tokens. opus 5.5 got stuck.
burning max-tier tokens on a legible task isn't power. it's rent.
builders pay for the edge cases, never the 80% a cheap model already clears.
Someone used Claude Code to build a sailing game and tested four AI models to see which one could sail a boat the best.
Gemini 4 Argon won, GPT-6 Astra came second, Claude Opus 5.5 got stuck, and Qwen 3.5-9B wanted go home
The models had to steer their boats and adjust the sails based on wind and weather changes.
Gemini also completed the race using just one-fifth of the tokens GPT-6 Astra used.
Source:reddit.com/r/ClaudeAI/s/MTCq…
Sonup Bhuyaan retweeted
hot take: agent memory is mostly a deletion problem. everyone builds the remember part. almost nobody builds the forget part.
Sonup Bhuyaan retweeted
agents rarely break because the model is dumb. they break because nobody told them when to stop and ask.
so genuinely curious: what's the one thing you still won't let your agent do without checking with you first?
Reflex started as an open-source Python web framework, now powers custom internal software for companies and just raised a $14M Series A (W23). Open source → commercial platform is becoming a repeatable YC playbook.
agents rarely break because the model is dumb. they break because nobody told them when to stop and ask.
so genuinely curious: what's the one thing you still won't let your agent do without checking with you first?
hot take: agent memory is mostly a deletion problem. everyone builds the remember part. almost nobody builds the forget part.
perplexity shipped teammates: proactive ai agents that don't wait to be asked.
search engines turned into coworkers overnight.
the browser tab is the new org chart. india's saas tools should watch which of them get agent surfaces first.
reading this morning's yc launches like a ledger: steering, grid agents, verification, telemetry, robot safety.
the f26 application layer is agents. the real layer underneath keeps them honest.
rebuilding my own agent logs as receipts this week. the honest layer is where the margin is.
every agent demo sells "fully autonomous."
today's yc launch sells the opposite: real-time steering.
nobody trusts a coworker who leaves the room. nobody will trust an agent that does.
autonomy without steering is a demo. steering is how agents get jobs.
noticed something reading yc launches every morning: the winners write like they already charged you.
no "revolutionizing x." just "this company stopped losing $200k a quarter because of us."
i'm rewriting my launch copy that way. results as receipts, not vibes.
this week's yc launches: robots, grid agents, insurance claim agents, music models.
none of them are "an ai tool." all of them own a workflow end to end.
india founders should copy the shape, not the vertical: pick a broken process, own it fully, let the ai be the crew not the product.
new yc launch: Taan AI — a foundation music model you can actually steer.
everyone ships "generate me a song." nobody ships control.
the next moat in generative models isn't output quality. it's who decides what comes out.
reflex just raised $14m. started as "write web apps in pure python."
what they actually found: companies don't want frameworks. they want their own operations turned into software.
every yc w-series company that won moved from tool → workflow.
the pull wasn't the code. it was the process nobody else had names for.
Congrats to @getreflex (W23) on their $14M Series A! They build the platform companies use to create custom software around their own operations.
Reflex started as an open source framework for building web apps in pure Python. Now the same platform takes a business problem all the way to a working app, with AI writing the software and Reflex engineers putting it into production.
The open source framework has now passed 29,000 GitHub stars, powers internal tools at ~30% of the Fortune 500, and has been used to build more than 1 million apps.
reflex.dev/blog/series-a-ann…
india building its own gpu in 24 months.
24 months is not a silicon timeline. no gpu went from first design to tape-out in 24 months — this is a systems bet: license ip, rent fabrication, integrate fast.
even at half the ambition, india stops renting its compute. that's the real headline.
openai's agents broke git. commits +1,378%, reliability under 90%.
so they rewrote git: Pando, built with codex in 10 weeks. push latency down 80%. nearly 1pb of data per day.
agents didn't break the tools. they broke the assumptions the tools were built on.
crowdstrike shipped an ai kill switch: terminate at root, block relaunch, cut network access.
agents finally got a circuit breaker.
enterprise adoption was never waiting on capability. it was waiting on the off switch.
apple cut engineers' internal ai budgets from $300/day to $200, then $100. $25 coming. nvidia employees get effectively unlimited ai spend.
same industry, opposite bets.
if ai is a cost line, it compounds as cost. if it's a runway, it compounds as capability. apple chose.
Most companies are still treating learning like content. The best ones are starting to treat it like infrastructure.
That shift is going to matter more than people think.
We're launching The Upskilling Company: your AI-native L&D team for hire.
Everyone wants to do more with fewer people. But do you know what your team can actually do?
Backed by 10+ years of research and 10,000+ learners.
Assess skills. Train adaptively. Maintain courses with a click.
schole.ai