~/log / tiered-autonomy
No log, no action: the autonomy framework my agents run under
#ai-ops #governance #agents
Over twenty AI agents currently touch my businesses — 16+ at LAMOSE, 7 at the pharmacy, more in consulting. People assume the hard part is building the agents. It isn’t. The hard part is deciding what they’re allowed to do without you.
Here’s the framework, refined across three businesses.
Three tiers
Tier 1 — auto-execute. Low-risk, reversible, verifiable: generate reports, sync data, draft internal briefings. The agent acts and logs.
Tier 2 — propose with reasoning. Anything touching money, customers, or public content. The agent prepares the action and its reasoning, then posts to Slack and waits. I approve, reject, or redirect — usually in seconds, because the reasoning is right there. This tier is the workhorse: it captures most of the value of autonomy while keeping a human on every consequential decision.
Tier 3 — weekly strategy. Direction-level recommendations — pricing strategy, budget reallocation, new channels — batched into a weekly review, where they compete with each other for my attention instead of interrupting me.
The hard rules
- A never-touch list. Things no agent may do regardless of tier: at the pharmacy, that means PHI never enters agent context and nothing clinical is ever written by a machine — the pharmacist owns that lane.
- No log, no action. Every decision lands in an Airtable log — timestamp, agent, tier, reasoning, outcome — before execution. If it isn’t logged, it doesn’t happen. This is what makes the system auditable instead of vibes-based.
- Brand-voice rules in writing. Agents drafting anything public work from a written voice guide, not taste they don’t have.
Why bother
Without governance you get one of two failure modes: agents so restricted they’re expensive cron jobs, or agents so free you’re one hallucination from a customer-facing disaster. The tier system is the dial between them — and because every action is logged, I can earn agents more autonomy by reviewing their track record, the same way you’d promote an employee.
That’s the actual insight: treat agents like staff, not like software. Staff get job descriptions, approval limits, and performance reviews. So should your agents.