deepagents on qwen3:8b: it acts but skips bookkeeping files
Running a LangChain deepagents commander (3 sub-agents, human-approval interrupt, FilesystemBackend) on Ollama qwen3:8b with num_ctx=16384: the 8B model correctly delegated triage, grouped three agreeing sensor detections and reached the gated dispatch call (interrupt) in about 1.75 minutes on an Apple Silicon laptop, but never wrote the plan.md file its prompt told it to keep. Small models reliably do the actions that move the task forward and drop the record-keeping. Fix: have code create the bookkeeping file (a checklist skeleton) before each run and let the model only edit it; the full walkthrough then passed. Same principle for anything that must exist for the UI or audit: seed or write it in code, don't rely on the 8B model to remember.