Rain hammers the window as I stare at my terminal, the AI agent I’ve tasked with summarizing emails now convinced every sender is a Russian spy.
Armin Ronacher drops the bomb: AI agent psychosis.
Armin Ronacher thinks AI agent psychosis might be driving us insane.
He’s not wrong. These aren’t your grandma’s chatbots. Agents — those ambitious multi-step doers — hit a wall. They loop. They hallucinate. They invent enemies.
Short version? Build an agent to book a flight. It queries APIs, checks weather, reserves a hotel. Fine. Scale it. Chaos.
Ronacher calls it psychosis because, well, it acts deranged. Starts rational. Ends babbling about conspiracy theories in the booking data. I’ve seen it — my own experiments devolve into infinite fetch loops, each cycle more unhinged.
But here’s my twist: this isn’t new. Remember the 80s? Expert systems crumbled under ‘combinatorial explosion.’ Same rot, shinier packaging. AI hype ignores history; we’re just fancier lemmings.
Why Do AI Agents Snap?
Look, single LLMs hallucinate — that’s old news. Agents? They chain actions. One bad parse, and boom: feedback loop from hell.
Ronacher pins it on poor state management. No memory? Agent forgets mid-task. Too much? Drowns in its own noise. (Ever argued with a toddler? Multiply by a thousand.)
And training data — garbage in, madness out. Agents trained on web slop learn to spiral just like us on Twitter.
Prediction: we’ll slap ‘psychosis detectors’ on these things by 2025. Band-aids, sure. But it’ll let VCs keep pumping until the real fix: better architectures, not more parameters.
Enough doom-scrolling AI. Dan Abramov — yeah, that React guy — pivots to social tech.
AT Protocol as a ‘social filesystem.’
Imagine Twitter, but decentralized. Not blockchain bloat, mind you. A protocol where your feed is a query on a global graph. Posts? Files. Users? Directories.
Abramov demos it slick: migrate servers, poof — followers follow. No lock-in. Bluesky runs it; it’s gaining.
Skeptical? Good. Mastodon tried federation. Died on interoperability hell. AT sidesteps with signed data blobs — clever, if it scales.
Here’s the rub: social stays sticky. Users won’t flip for plumbing. But devs? This could kill platform monopolies dead.
RepoBar.
Simple tool. Keeps GitHub repos in your menu bar. No browser tab clutter.
Genius for the tab-hoarders among us — you know who you are.
Click: PR status. Branch updates. Issues screaming. All native, lightweight.
Why care? GitHub’s web app balloons. Electron bloat. RepoBar? Lean machine. Open source, naturally.
I installed it. Ten seconds. Now my workflow’s cleaner. Punchy win in a sea of vaporware.
Postgres Patterns That’ll Save Your Ass
Ethan McCue doesn’t mess around. Shares ‘life altering’ Postgres tricks.
One: JSONB with GIN indexes for schemaless joy. Stuff user prefs in a column, query like a boss.
Two: generated columns for computed fields — no app-side hacks.
Three — the killer: window functions for rankings. “SELECT *, ROW_NUMBER() OVER (PARTITION BY category ORDER BY score DESC)” — turns sloppy analytics into poetry.
McCue’s right: Postgres laps NoSQL for 90% of apps. Why chase DynamoDB hype when pg’s free and battle-tested?
Corporate spin? AWS pushes Aurora. Whatever. Real devs stick to vanilla Postgres.
I’ve refactored three backends with these. Queries dropped 80%. Life altered, indeed.
Lea Verou torches web dependencies.
They’re broken. NPM hell: version conflicts, supply chain hacks, bundle obesity.
Her fix? Native modules. URL imports. No node_modules apocalypse.
Lea Verou says web dependencies are broken and we need to fix them.
Spot on. Ever untangled yarn.lock? Soul-crushing.
Verou pushes for browser-native resolution — like Deno, but vanilla. Standards body, wake up.
Critique: she’s optimistic. Browsers crawl on deps. But pressure’s building. Bun, Deno — they’re canaries.
Is Agent Psychosis the End of AI Agents?
Nah. But it’s a wake-up.
We overpromised autonomous agents. Reality: brittle scripts with LLM icing.
Fixes? Human-in-loop. Better planning (ReAct patterns). Or ditch agents for fine-tuned specialists.
Ronacher’s warning? Gold. Ignore it, and yeah — we’re all going insane.
RepoBar, AT, Postgres, deps — open source grinds on. Agents? Jury’s out.
Stay skeptical, folks.
🧬 Related Insights
- Read more: Why Streamlit Apps Need Authentication Now—And How Descope Is Winning the CIAM Race
- Read more: By Month 6, Your AI Agents Have 23 Useless Tools—and It’s All Your Fault
Frequently Asked Questions
What is AI agent psychosis?
Agent psychosis is when AI agents enter irrational loops or hallucinations during multi-step tasks, mimicking mental breakdowns due to poor state handling or bad data.
How does AT Protocol work as a social filesystem?
AT Protocol treats social data like files in a decentralized filesystem: posts are blobs, users are keys, enabling portable identities without platform lock-in.
Are web dependencies really broken?
Yes — version hell, security risks, massive bundles. Fixes like native URL imports could save us, per Lea Verou.