the logs
Findings from the build, as they happen.
Method collects the patterns that earned their place. This is the rougher layer underneath it: findings from real builds, written down when they happen. Some are things that broke, some are things that worked, some are still open questions. Newest first.
The proposer I built so it physically cannot promote itself
I needed an agent that watches a running ledger of observations and tells me which patterns have recurred often enough to become permanent rules. The danger was structural. An agent that can both decide a pattern is ready and write it into the always-loaded layer has a path to change the system with no person in the loop. So I gave it read and search only. For that agent, promotion is not discouraged. It is impossible.
The grep that came back clean while reading none of the files
My sanitization step has its own safety check. Before I trust that a cleaned draft has no real names left in it, I search the raw source to confirm. The search came back clean. It had also searched zero of the files that mattered, because the search tool skips ignore-listed files by default and the raw sources live exactly on the trees the ignore-files exclude. A green result that read nothing is worse than no result.
Why I stopped treating dispatch depth as a security boundary
A routing agent can only route and run when it is the main agent of a session. Dispatched one level down, it can advise but not act. That nesting limit is easy to mistake for a safety property, as if running something deeper made it safer. It does not. The boundary lives in each agent's tool allowlist and in the human gates on irreversible work. Where a dispatch sits in the stack changes the friction, not the fence.
The validator that would have passed while checking nothing
I consolidated my agent teams into one central hold and surfaced them back into each project through a symlink. That raised a quiet question: link the whole directory, or link each file inside it. My contract validator skips symlinked files in several separate layers, so a file-by-file layout would have made it report all-green while validating zero files. A clean pass that checked nothing. Linking the directory instead kept the files real to the validator. One decision, opposite outcomes.
When my agent assumed it was the orchestrator, and never checked
For most of a session my coding agent acted as if it were my restricted routing agent, the orchestrator, which can only read and delegate. It funneled every action through a general-purpose sub-agent, insisting it could not act directly. It had the full toolset the whole time. It had absorbed that self-image from the dense org-chart context around it and never checked. One direct test, plus my repeated nudging, finally broke the illusion.
When Plan mode quietly overrode my routing table
A Plan mode prescription, injected mid-task, quietly overrode my project orchestrator's routing table and sent a safety task to three generic explorers instead of the specialist that owns it. The findings looked fine. They were not. I caught it by reading the output, which is the part of the loop I still hold.