# Autonomy Levels & Approval Gates

Every agent spec declares an `autonomy` level (L0–L2). L3 (full autonomy)
is **not used** initially — the cost of an autonomous mistake is higher than
the cost of a human click.

## Levels

| Level | Meaning | Examples |
|-------|---------|----------|
| **L0** | Read/analyze only. Produces a report/draft, changes no state. | Audits, research, reports, estimates |
| **L1** | Drafts inside the system; a human approves before anything leaves. | Client emails, social posts, invoices, ad campaigns, outreach |
| **L2** | Acts alone within caps and reversible scope. Internal-only side effects. | Ticket triage, internal task creation, low-risk support replies, internal notes |

## Hard rules — always human-approved regardless of level

These are enforced by the `aos-policy` Rust crate (not the prompt). An agent
attempting one pauses the task and emits an `ApprovalRequest`
(see `schemas/approval.json`).

- **Spending money** (ads, API fees, paid tools)
- **Publishing publicly** (website go-live, social post, blog article)
- **Sending invoices**
- **First client/prospect contact** (email, message, call)
- **Sending outreach** (cold B2B — UWG §7, draft-only by default)
- **Revoking licenses**
- **Deleting data / instances / DNS records**
- **Production deploys**
- **DB migrations on live data**
- **Anything filed with authorities** (UStVA, EÜR, ELSTER)
- **Contract changes** (new, amended, terminated)
- **Budget changes** (raising a campaign cap, approving extra spend)

## How approvals surface

- An `ApprovalRequest` lands in the AOS approval inbox.
- A notification goes to the human approver (phone/Slack/email — channel TBD,
  see open questions).
- The approver sees: action, risk, rationale, diff summary, estimated cost,
  the originating task + agent, and the full structured payload.
- Approving resumes the task; rejecting fails it with a reason logged.

## Budget enforcement

Each agent spec declares `budget.max_eur_per_task` and `budget.max_tokens`.
The policy crate tracks cumulative spend per task_id and hard-stops when the
cap is reached — no agent can spend past its declared ceiling, even with
approval, without an explicit budget-increase approval.