Spend human attention on ambiguity
spend-human-attention-on-ambiguity
Summary
Agent throughput can exceed the maintainers ability to review and steer. Delegate execution and inspection, but reserve human attention for ambiguous goals, taste, architecture, risk, and the systems that make future delegation safer.
Problem
Agent throughput can create more work than the maintainer can review. Human attention is still the scarce resource, especially for ambiguous goals, product judgment, new architecture, and decisions where the acceptance criteria are not yet known.
Preferred move
Use agents for execution, inspection, mechanical cleanup, and well-specified changes. Spend human attention on choosing goals, resolving ambiguity, setting taste and constraints, reviewing high-risk outputs, and improving the systems that make future agent work safer.
Carry the user's existing authorization forward. Choose routine details, representative fixtures, and reversible local checks inside that scope without asking the user to restate the request. If a repository default appears to conflict with a later explicit request, check whether the request already supplies the exception. Explain any remaining constraint precisely.
Ask when an unresolved choice materially changes the goal, contract, risk, or authorized scope. Continue independent work while that choice is pending. Prepare the concrete artifact and evidence before requesting approval for publication or another action that still needs it.
Tradeoff
Do not abdicate judgment just because a task is easy to delegate. If the output changes product direction, security posture, public API, or long-term architecture, keep a human decision point in the loop.
Agent instruction
Before doing broad work, identify which decisions need human attention and which parts are safe to execute within existing authorization. Resolve routine details yourself; ask about material ambiguity before burying it in a large diff.Examples
Bad: the agent consumes review attention on a large ambiguous bundle.
I redesigned the onboarding flow, changed the data model, and updated the API because it seemed
cleaner.
Good: the agent separates execution from the human decision.
I found three onboarding data-model options and recommend option two. Once chosen, the migration and
tests are mechanical.
References
- OpenAI harness engineering: frames human attention around goals, environment design, and feedback. Existing authorization still determines which decisions an agent can make in a particular task.