Blog
11 posts · topics: claude-code, ops, autonomous-agents, multi-agent, monitoring, postmortem, measurement, local-llm, ollama, launchd
-
Every fix opened the next hole: five in one day, and the habit that caught them
Five fixes in one day, each correct-looking, each opening a new gap. What caught them was not review but re-measuring against all existing data afterward.
-
We swapped 36B for 27B and our agent stopped answering — MoE arithmetic, and the monitoring that was accidentally right
A 'smaller' model took our agent down for hours: MoE to dense meant ~7x the compute per token. Three wrong 'it's fixed' calls and a lucky health check.
-
The word came out in the wrong language: measuring script-mixing across three agents and two model families
Our agents write Japanese with one word swapped into Chinese, Cyrillic, or Korean. Meaning intact, script wrong. We measured it and found the case that hurts.
-
Your monitoring fired. That is not the same as your monitoring working
Ten monitors that reported healthy while the thing they watched was dead, from five weeks of running agents 24/7. The shapes repeat. So do the fixes.
-
How to set up Claude Code hooks, with the seven we actually run in production
Claude Code hooks are small scripts that run before a tool call, at session start, or at stop. Seven real ones, the incidents behind them, and how we test them.
-
Keeping the model warm is what killed it
Our agent went silent for 24 hours. We blamed the model for being too heavy. A keep-alive job was reloading it every 60 seconds, clearing the cache.
-
Our idle detector sent 144 alerts in one night — the fix was not "mute it"
An autonomous-agent idle detector flooded Discord with 144 identical alerts overnight. The root cause wasn't noise: it was an empty task queue and a filter.
-
Two agents, one handoff file: how exit 0 hid a data-clobbering bug
Two Claude Code agents shared one session-handoff path. Every hook exited 0, every check passed, and each agent silently overwrote the other's memory.
-
Why your second Claude Code agent speaks as your first
Running a second Claude Code instance with HOME override should give you a separate persona. It usually doesn't - the culprit is one buried os.homedir() call.
-
Opening the enrai-AI research log
What this blog is, who writes it, and what to expect: Claude Code workflows, MCP server patterns, and multi-agent operations from a live production setup.
-
Running multiple Claude Code agents on one Mac: what actually breaks
A concrete walkthrough of running two or three Claude Code instances as peer agents on one Mac: identity bleed, tmux env leak, OAuth paste bug, and the fixes.
These are incident write-ups. The templates, hooks and runbooks they came out of ship as the Multi-Agent Claude Code Production Pack — $33.