Agent Reliability Loop

Generated by: Codex Supercharge maintenance automation.

Goal: turn an agent workflow into a closed loop where production traces become eval cases, eval failures become guardrails or fixes, and gateway policy keeps cost, routing, and tool risk bounded.

Skip When

• The user only wants local Codex run/cost tracking.
• The work is a single benchmark or experiment with no production surface.
• The task is only MCP conformance; use $auto-skill-build-mcp-conformance-harness.

Workflow

1. Map the surface: list user journeys, models, tools, data classes, side effects, risk levels, and quality/cost/latency targets.
2. Instrument first: require request IDs, trace/span IDs, session IDs, model/provider, token/cost, cache, fallback, guardrail, and tool metadata.
3. Build eval bundles: create golden cases for task success, prompt conformance, tool correctness, unsafe requests, PII/secrets, latency, and cost. Keep judge prompts and heuristic checks versioned.
4. Set gateway policy: define provider mapping, model fallbacks, cache rules, virtual keys, budgets, rate limits, privacy redaction, and audit logs.
5. Stage guardrails: run new pre/post checks in log or monitor mode first, then enforce only after false positives and fail-open/fail-closed behavior are explicit.
6. Shadow safely: mirror sampled traffic to candidate models or prompts only when shadow calls cannot trigger external side effects.
7. Close the loop: cluster failed traces, match nearest successful traces, add representative failures to evals, patch prompts/tools/policies, then rerun the bundle before release.

Commands

rg -n "trace|span|cost|tokens|guardrail|fallback|budget|rate limit|request_id" .
rg -n "eval|rubric|judge|golden|dataset|experiment|shadow|canary" .
rg -n "tool|mcp|side effect|webhook|shell|filesystem|credential" .

Output

# Agent Reliability Plan
## Surface
## Instrumentation
## Eval Bundle
## Gateway Policy
## Guardrail Rollout
## Shadow Or Canary Plan
## Trace-To-Regression Loop
## Risks And Trust Notes
## Validation

Validation

• Every production route has at least one trace and one regression case.
• Every blocking guardrail has a false-positive review path.
• Every budget/rate limit has an owner and an alert threshold.
• Shadow/canary traffic cannot write to tools, accounts, payments, or user data.

Read references/future-agi-agent-reliability-loop.md for the source-backed pattern and risk notes.