Automate the useful parts, govern the risky parts

Lesson 05.1

Do you need an agent?

A fixed workflow is usually easier to test, explain and maintain. Use agentic behaviour when a task really requires adaptive choices about tools or next steps.

Before granting more autonomy, define who owns the process, what sources and actions are permitted, which changes are reversible, and how errors will be detected.

Lesson 05.2

An action ladder

This ladder is a planning heuristic, not a certification. Autonomy is a risk decision, not a reward for an impressive demonstration.

• Suggest — AI recommends, a human decides.
• Draft — AI prepares work, a human reviews before use.
• Act with approval — AI can execute a specified action only after approval.
• Limited autonomous action — AI acts in a bounded, reversible domain with monitoring.
• Expanded autonomy — broader scope only after demonstrated reliability and stronger controls.

Lesson 05.3

Design the unhappy path first

For sensitive operations, use permissions and external approval controls rather than relying on a model promise not to do something. Keep logs proportionate and privacy-aware. Use duplicate prevention or idempotency where appropriate.

• What if a tool is unavailable or returns only part of a result?
• What if the sources contradict each other?
• What if the workflow runs twice?
• What if the result would publish, spend, delete or contact someone?
• What if the system cannot verify that an action succeeded?
• What is logged, who is notified and how is the action rolled back?

Trigger → retrieve and validate → decide → act or request approval → record and monitor

For every arrow, add: failure behaviour, owner, timeout or retry policy, and a way to detect that it did not complete.