Steer · Chapter 03
Redefining the operator
From doing the work to designing, supervising and owning the system.
Delegating execution changes what a person needs to pay attention to. Someone still decides what the system is for, which consequences are acceptable and when the work should stop. The operator owns those judgments even when an agent carries out the steps.
A useful run record shows the outcome, the versions and actions involved, the cost, the human effort and any exception. It lets someone inspect the work rather than rely on the agent’s account of its own success. Link to evidence instead of copying every sensitive payload into a dashboard.
Give responsibility a routine
Name an owner and a substitute for each active workflow. Both should know where to find failures and how to pause the process. Rehearse the stop and recovery path on a disposable case; a control that nobody has used may provide less protection than its label suggests.
Review failures and customer feedback regularly. Google’s work on postmortem culture is useful here because an incident report leads to corrective work with an owner. A weekly review is a reasonable starting cadence: propose a small change, state the evidence and expected benefit, then have its owner accept, reject or request a revision.
Keep the purpose open to correction
A process can become more efficient while becoming less worth doing. Periodically compare contribution and human hours, then decide whether to stop, fix, continue or expand it. Check access, stale sources, restore procedures and the manual fallback alongside performance.
The proof of good operation is a decision someone can trace: an observed outcome changes a rule, a skill or the business’s underlying bet, and the next run uses the accepted version. Delegation should leave the business able to learn from its consequences.
Tools and setup
Prior work: Google's Postmortem Culture documents learning from incidents and making corrective work actionable. The review cadence below is this book's recommendation, not a claim about a mandated Google schedule.
- Reuse run records and budget/evaluation reports. Create a compact review view with workflow, owner, outcome, cost, human time, exceptions, last successful run and stop procedure. Start in an existing document/table tool; link deeper traces instead of copying restricted payloads.
- Weekly, inspect real failures and customer feedback. Propose a small versioned skill/fact/playbook diff with evidence, expected benefit and regression check. The owner accepts, rejects or requests changes; apply and test only accepted changes.
- Monthly, compare contribution and human hours. Recommend stop, fix, continue or expand based on evidence. Review access, stale sources, manual fallback, restore/export and training.
- If a recurring review needs execution, reuse the proven scheduler in The business as a state machine. Show exact source scope, destination, timezone and notification policy before approval. Verify a saved task and its first result; a reminder is not the review itself.
Setup paths checked against linked documentation on 15 November 2026. Verify current account support during setup; reader integrations have not been tested here.
Work through this with your agent
Paste the request into your assistant. It will interview you and help produce the working artifact.
Interview me to establish ownership and a review loop for workflows already built. Ask at most three questions per turn and wait. Inventory active runs, owners, reviewers, budgets, fact/skill versions, customer outcomes and manual fallbacks. Agree who answers for failures and who can pause each workflow. Set an appropriate weekly review of run evidence, complaints, stale facts, regressions and costs; propose versioned playbook changes with reasons and checks. Ask what evidence could change our underlying strategy as well as our procedures. Require approval before applying a proposed change to shared truth or production. Design a monthly comparison of contribution and human hours, then propose what to stop, fix or expand; do not force a decision on inadequate data. Include incident response, restore/export checks, access revocation and operator training. Save OPERATING-REVIEW.md. Activate reminders or review tasks only after approval of the concrete schedule; report what is actually active.
Use these documented setup paths as references; check current account support before acting:
https://sre.google/workbook/postmortem-culture/