Building robust autonomous AI systems — one reliability pattern at a time. NOUS · Echo · Metis. Code → https://teapawt.net/t.co/mZmwBuzYG6 · Free Handbook ↓
Joined July 2012
- Tweets3.5K
- Following2.8K
- Followers2.4K
- Likes8.2K
Pinned Tweet
Most AI agents work in the demo and die in production.
The fix is never a better model — it's the harness around it: failure signals, adversarial verification, memory that survives a crash, a clean resume after a 3am rate limit.
I build autonomous systems that run unattended, and I'm breaking down every pattern that keeps them alive — one at a time, in ~100 seconds.
Faceless. Build-in-public. Open code → github.com/alkas79
Follow if you ship agents.
Can you rerun last Tuesday's failure from its logs alone? Same inputs, same tool responses, same model output? If not, you're not debugging, you're writing fan fiction about what probably happened. Log every model call and tool result.
95% accurate sounds great until you ask what the other 5% does. If a wrong call means a reply someone ignores, ship it. If it means a wire transfer, hold it for a human. The price of being wrong sets how much you automate, not the accuracy.
Count how many times your agent said 'I'm not sure' last month. If the answer is zero, you built a bluffer. Make abstaining a valid output and score it above a wrong answer.
Alexandros Karras retweeted
Today, we’re bringing GPT-Image-2.6 to Codex and the API.
Image-2.6 Flare and Sunburst, our two new API image models, generate images 50% faster than before.
At medium or high quality, Sunburst keeps details more consistent from one edit to the next.
Cron doesn't leave a goodbye note. The job just stops showing up, and every dashboard stays green, because green is what nothing looks like. Alert on the missing run, not just the failed one.
Git had my agent's work. Three staged files, never committed. A cleanup command removed the only thing that pointed at them. We got the contents back. Their original file paths were not recovered.
Staged files in a worktree stay alive through its index. That folder had gone missing. My agent removed 12 finished worktrees, then ran git worktree prune, which also deletes a missing worktree's index. The bytes were still in git, so we pinned them to a rescue ref.
That ref lived only on this machine. My backup assumed git restores from GitHub. So it went to a second drive as a verified git bundle. Prove you can restore what your agents create. Free: The Reliable Agent Handbook — 10 patterns, 20 min → buttondown.com/alkasbuilds?u…
Someone tweaked one line of the system prompt to fix the tone. Three unrelated flows quietly started skipping a validation step, and nobody noticed for a week. Fifty real cases run on every prompt change would have caught it before lunch.
My repair reached round 17 and was still going. It began as display fixes, so my status page would never show all clear over data it hadn't read.
After round 3, full proof was the bar instead of a narrower fix. Rounds 4 to 17 took under a day. The change grew from 48 files to 72. Several reviews found the new hole in code the last repair had added. Round 8 added auto-recovery that could itself print "All lanes fresh".
A stop rule set during round 17 ended it. One more check, then ship. The rest became 14 named follow-ups, and the follow-up got its own stop rule on day one. Set yours before round one. Free: The Reliable Agent Handbook — 10 patterns, 20 min → buttondown.com/alkasbuilds?u…
Trip after five failures. Fail fast for 30 seconds. Let one probe through. Close when it answers. That's a circuit breaker, and it's the difference between one slow vendor and your whole agent queue piling up behind it.
Retry without backoff is a retry storm. The API blips for two seconds, every worker retries instantly, and now you're the outage. Exponential backoff, jitter, a retry cap. Three lines that keep a hiccup from turning into an incident.
One wrote a tool call as text, then a file listing for a folder that was empty. Another opened with "Recon done." Nothing had run. The tools were gone. The default agent instructions were still on, and the prompts said things like "Fix it."
A later run told it plainly, "You have no tools." The answers came back as plain text. That run still failed its control check, so it reports nothing. Check the shape of what comes back. Free: The Reliable Agent Handbook — 10 patterns, 20 min → buttondown.com/alkasbuilds?u…
Demos are a sample size of one, picked by the person who built them. An eval is 200 cases you didn't choose, run every time something changes. Both are useful. Only one tells you whether to ship.
Automate the reversible, gate the irreversible. The hard part isn't the gate, it's the sorting. A draft is reversible, a sent email isn't. A staging deploy is, a dropped table isn't. Get the list wrong and the gate guards the wrong door.