Field notes · AI agents in production
What breaks when agents ship
I build multi-agent systems and write down what went wrong. Guardrails that reported safe while switched off. Probes that could not run. Gates that were never wired to anything. Everything here is something I hit myself — not a tutorial, and not a summary of someone else's postmortem.
Silent failure
A check that cannot run and a check that found nothing produce the same output. Most harnesses cannot tell them apart.
Failure direction
An absent gate is survivable. A gate that is absent and still reports “safe” is worse than no gate at all.
Half-done migrations
Changing the config and the docs is not changing the system. Something is still launching the old thing.
Latest
All writing →A probe that cannot run looks exactly like a probe that found nothing
A check that silently fails open will cause the very regression it exists to prevent. Here is how I nearly shipped one.
Read →This is a new site and there is not much here yet. More is being written — the RSS feed is the way to catch it.