Practice Agent Workouts
practice-agent-workouts
Summary
High-risk or repeated agent workflows are unreliable when first exercised during real work. Build safe practice tasks with explicit win conditions so the workflow, skill, or runbook can improve before pressure arrives.
Problem
High-level workflows are unreliable when they are only tested during real work. Login flows, uploads, permission changes, UI navigation, and deployment procedures need practice targets before they can be trusted under pressure.
Preferred Move
Create practice tasks with explicit win conditions. Use safe fixtures, flags, canaries, or sandbox states so the agent can exercise the workflow, reflect on failures, update the skill or runbook, and rerun until reliability improves.
Tradeoff
Do not build elaborate training harnesses for one-off tasks. Use workouts for workflows that will be repeated, delegated, or allowed to run with less supervision.
Agent Instruction
When a workflow is important and repeatable, propose a safe workout that exercises it end to end.
Capture failures as updates to the skill, runbook, or validation checklist.Examples
Bad: the first real permission change is the first full test of the workflow.
Grant the production group access and see whether the UI flow works.
Good: the agent practices against a harmless target first.
Grant access to the sandbox group, verify the expected marker appears, update the runbook for the
two-factor prompt, then repeat before touching production.