Optimization Pattern — MCP recommendations
Pattern: Optimization
Component: mcps.md
Version: 1.1 | Updated: 2026-07-29
Purpose
Least-privilege MCP servers for the Optimization pattern. Keep mcpServers: {} until a human enables stubs after MCP when and how.
Recommended MCPs
| MCP | Why for this pattern |
|---|---|
| Datadog | Latency/error baselines |
| Grafana | Optional dashboards |
| Sentry | Error regression watch |
When to enable for Optimization
Datadog/Grafana/Sentry provide baselines and regression watch. Enable them before changing hot paths; keep write/deploy tools off.
Enable rules (this pattern)
- Optimization MCPs are measurement planes — never auto-tune production
- Never commit tokens; map secrets outside git
- Record allowlist + data class in MODEL_POLICY / SoA before write tools
- Freeze dashboard queries that define the baseline so they cannot silently change mid-run
Anti-patterns
- Optimizing against a dashboard that was edited during the experiment
- Shipping a hot-path change because p50 improved while p99 was ignored
- Using MCP to mutate feature flags without a correctness oracle
Checklist
- Optimization MCPs mapped to env secrets outside git
- Unused servers remain disabled after the optimization gate
- Injection / tool PEP reviewed if tools can act
Version: AIES v1.0.0✏️ Edit this page on GitHub