Build an autonomy ladder so your agent knows when to ask
Yesterday I talked about the "fix first, report after" pattern. The immediate question is: fix what first? How does the agent know the difference between "restart a crashed service" (just do it) and "reply to a legal threat" (definitely ask)?
The answer is an autonomy ladder — a tiered system that maps every category of action to a level of independence. Think of it as your agent's decision-making constitution.
Tier 1: Act immediately, report after
These are actions where speed matters more than your input, and the downside of acting wrong is low:
- Fix monitoring alerts and restart crashed services
- Answer routine customer questions (where's my order, how do I reset my password)
- Update internal documentation
- Rotate expired credentials
- Run scheduled reports
Tier 2: Act immediately, report with detail
Higher stakes, but still within your agent's competence. The key difference is the reporting — you want enough detail to audit the decision after the fact:
- Process refunds under $50
- Deploy bug fixes to production
- Send follow-up emails to leads or customers
- Triage and respond to support tickets
- Make minor content updates
Tier 3: Propose and wait for approval
These are actions where the agent should do all the prep work — research, draft, recommend — but wait for your green light before executing:
- Any financial commitment over $100
- Outbound communication to press, legal, or investors
- Architecture or infrastructure changes
- Anything involving unreleased products or sensitive data
- Hiring, firing, or contract decisions
Put this in your MEMORY.md and reference it from SOUL.md. The agent internalizes these boundaries and stops asking permission for Tier 1 work while still escalating Tier 3 decisions.
The liberating thing about the autonomy ladder: once it's defined, you can expand it over time. As you build trust, you promote actions from Tier 3 to Tier 2, or from Tier 2 to Tier 1. The ladder grows with the relationship.
Start conservative. Let your agent earn autonomy. But start — because an agent that asks permission for everything is barely more useful than a to-do list.