Designing Human and AI Agent Work
Divide human and agent work using consequence, ambiguity, verifiability, context, reversibility, and accountability as design criteria.
Decompose the workflow
Break work into goals, decisions, actions, inputs, outputs, and exceptions. Whole jobs are usually too coarse a unit; specific tasks reveal where autonomy is safe and where human context remains essential.
Set autonomy by consequence
Grant more autonomy when outputs are testable, actions are reversible, and boundaries are clear. Require review when errors affect rights, money, safety, commitments, sensitive data, or important relationships.
Design the learning loop
Capture corrections, exceptions, and outcome feedback in a form that improves prompts, policies, tools, data, and training. Do not rely on individual employees to compensate indefinitely for weak system design.
Common Mistakes
- Automating entire roles at once
- Using human review as an undefined safety net
- Failing to log agent actions
Market Signals
- Agents receive broad goals without tool limits
- Reviewers approve outputs without useful context
- The same corrections recur without system changes