Garbage Collect Agent Drift
garbage-collect-agent-drift
Summary
Agent-generated notes, plans, and assumptions can outlive their usefulness and mislead later work. Periodically remove, update, or promote them so durable guidance stays current and temporary scaffolding does not become false context.
Problem
Agent-generated systems accumulate drift by copying existing uneven patterns, spreading weak abstractions, and leaving stale docs or tests behind. Waiting for a large cleanup lets the drift compound.
Preferred Move
Run small recurring cleanup loops. Encode golden principles as patterns, lints, tests, quality grades, or focused review checks. Let agents open targeted fixups that are easy to review and merge.
Tradeoff
Do not churn code just to make it uniform. Garbage collection should remove real drift: stale docs, broken boundaries, repeated weak patterns, noisy tests, or quality gaps that will mislead future work.
Agent Instruction
When you find repeated drift, propose the recurring check or small cleanup path. Prefer focused
fixups that keep future agent runs aligned.Examples
Bad: cleanup is saved for a painful rewrite.
We should eventually clean up all the inconsistent schema helpers.
Good: cleanup is continuous and scoped.
Added a check for schema helper naming and opened a focused follow-up to normalize the three
violations it found.
References
| Source | Use | Note |
|---|---|---|
| OpenAI garbage collection | supports |
Agent systems need recurring cleanup to prevent entropy and drift. |