Observe
Collect representative work and label the decisions you make, including rejected cases.
Turn the judgment in your head into an agent runbook.
You already know why the clean-looking SOP fails on Tuesday afternoon. Operator work starts with those exceptions, then builds instructions, review thresholds, and recovery paths an agent can follow without hiding uncertainty.
Extract the real decision system
What the SOP says
- Check the request
- Update the system
- Notify the owner
- Close the task
What the agent needs
- Which source wins when records disagree
- What counts as enough evidence
- When a request is urgent, risky, or incomplete
- Who approves an irreversible action
Release automation in four lanes
Draft
Let the agent recommend an action while the operator remains the only person who can execute it.
Shadow
Run agent and operator in parallel. Track disagreement by category, not just an average score.
Delegate
Automate the low-risk lane, preserve the review queue, and keep every decision inspectable.
Operator artifact
The exception ledger
A durable ledger turns edge cases into better instructions instead of one-off corrections that disappear in chat history.
- Trigger
- The input or condition that exposed the gap.
- Wrong action
- What the agent proposed or executed, with sensitive details removed.
- Rule change
- The instruction, threshold, or tool boundary that changed.
- Regression case
- A saved example that must pass before the next release.
Bring a recurring workflow and its ugliest exception.
Agent Ops is the paid working room for logs, blueprints, and monthly review. Builder is the structured curriculum when you need the complete system from instructions through deployment.