# Cost Controls And Runtime Budgets Review Checklist

Use this checklist before launching loops, retries, retrieval, delegation, write tools, or multi-agent work.

## Budget Policy

- [ ] Every run starts with task class, risk class, user tier, and budget policy version.
- [ ] Budgets cover cost, model calls, tool calls, write tool calls, retrieval, delegation, retries, wall-clock time, and approval.
- [ ] Budget policy differs by task value and risk instead of using one global limit.
- [ ] Budget changes are versioned and reviewable.

## Enforcement

- [ ] Budgets are checked before model, tool, retrieval, memory, delegation, and approval actions.
- [ ] Projected cost is checked before the next action, not only after spend occurs.
- [ ] Runtime-owned counters are the source of truth.
- [ ] Tool wrappers cannot bypass budget enforcement.

## Degraded Modes

- [ ] Low-token behavior preserves evidence references and summarizes state.
- [ ] Low-model-call behavior stops revision loops and returns the best validated result.
- [ ] Low-tool-call behavior moves from action to draft or read-only analysis.
- [ ] Low-wall-clock behavior returns partial results or queues background work.
- [ ] Low-retry behavior stops retrying and exposes the blocker.
- [ ] Low-delegation behavior assigns one owner or stops.

## Observability And Evals

- [ ] Traces record starting budget, remaining budget, budget decisions, degraded mode, and stop reason.
- [ ] Operators can explain cost for a single run.
- [ ] Evals cover exhausted model calls, write-tool limits, approval thresholds, delegation limits, wall-clock limits, and low-value expensive routes.
- [ ] Final answers fail eval when they exceed budget without approval.

## Operations

- [ ] High-risk runs can request approval for more budget.
- [ ] Low-risk runs can stop without human review.
- [ ] Budget policy can be rolled back.
- [ ] Incidents caused by budget policy can be replayed against old and new policy versions.