Your agents did 300 things last night.
Which five need you?

agent-steward is a quality gate and spend auditor for AI agent fleets: it checks their output against your own rules, meters every token from the transcripts your harness already writes, and shows you only what genuinely needs human judgment.

$ pipx install agent-steward

GitHub PyPI Reference MIT · zero network by default · read-only on your repo

The problem

When one agent does one task, you review it yourself. When a fleet does hundreds a day, you can't — so you either review everything (you're the bottleneck), review nothing (silent errors compound), or cap spending with a blunt dollar limit that can't tell money well spent from money wasted.

There's a fourth option: machines check what machines can check, you judge only what needs judgment, and every dollar gets a task-level justification.

Measured, not promised

68%
estimated spend saved vs running everything on the top model — escaped-defect rate unchanged (0.0009)
289 → 5
a week of flagged items compressed into the decisions that actually needed a human
4×
real spend vs what hand-logged estimates had recorded — measured straight from session transcripts
6 min
to onboard two projects the tool had never seen, using zero new check types

Every number above is reproducible from the ledgers the tool writes. Teal does the checking; amber is the one thing that needs you.

One loop, four verbs

Verify — your written rules become deterministic checks that run after every agent session → Allocate — every task class gets a model tier; failures escalate up, canary shadow-runs prove where cheaper holds quality → Fix — findings arrive grouped by category, each row stating how it gets fixed and who may fix it; you authorize in one sentence → Judge less — everything the loop learns shrinks what reaches you next round. Cheapest at the same quality, never cheapest at any quality.

Who does what

stewardChecks, meters, sorts, and proposes — tier changes, rule candidates, an ordering of your attention. It never edits your files, never decides, never phones home.
your agentDrafts the rules from your own docs, fixes only the categories you authorize, self-repairs new violations the hook feeds back after each session.
youRead one section — “What needs you” — and make the calls no machine may make: rule conflicts, tier floors, quality verdicts. Most days that's under five items.

Set it up with one prompt

Open an agent session inside your project and paste:

Set up agent-steward for this project:
1. Install it: pipx install agent-steward
2. Run `steward init --out steward.yaml`, read the rubric it prints, and fill in
   the skeleton from THIS project's own written rules (CLAUDE.md, docs/, schemas).
3. Show me the drafted steward.yaml — walk me through only the decisions that
   are mine (severities, tier floors).
4. After I approve: steward baseline, steward install-hook, steward ingest-usage.
5. Run `steward report` and explain the "What needs you" section to me.
# from then on: the hook runs after every session, `steward report` is your dashboard