## Definition
**Decommission discipline** is the practice of actively retiring automation — agents, skills, hooks, MCP servers, standing rules — that no longer earns its place in the platform. The under-discussed counterpart to building.
## Symptoms an Agent Has Overstayed Its Welcome
- **Humans always override its verdict.** A critic overruled 80% of the time is a tax, not a filter.
- **It never finds anything.** A security auditor that has passed every PR for two months is either perfect or asleep. Test it; bet on the latter.
- **It introduces a new failure mode.** A release manager that occasionally publishes broken versions is worse than no release manager.
- **It's a single point of dependency on a brittle prompt.** When the underlying model changes, brittle prompts break.
## The Cure Is Rarely "Delete the Agent"
It is usually: **tighten the spec the agent works against**, then re-evaluate. See [[Spec-Driven Development]].
## Decommission Audit (quarterly)
For each component installed (agents, skills, MCP servers, hooks, `AGENTS.md` rules):
| Component | What it does | Last earned its keep | Verdict |
| --------- | ------------ | -------------------- | ------- |
- KEEP — has a recent, concrete event where it added value.
- TUNE — is doing the wrong thing or producing too many false positives.
- RETIRE — no recent value or actively harmful.
## Why It Matters
- Untended automation accumulates as a tax on every session — tokens, false positives, cognitive overhead.
- Older prompts were often built around model behaviours that newer models have surpassed; the underlying gap has closed but the cruft remains.
- A clean platform is also a measurable one ([[DORA Metrics]]).
## Hard Move
Actually delete components marked RETIRE. *No commented-out leftovers.* They rot.
## Related
- [[Orchestrator Role]]
- [[DORA Metrics]]
- [[Specialized Agent]]
- [[Hook]]