The short version

Feedback is most valuable when it improves future work, not only the output in front of a reviewer. Anthropic’s account of Warp describes a useful pattern: separate the agent doing the task from an improvement process that studies accumulated feedback and proposes focused changes to reusable instructions.

The proposal remains reviewable. The system does not silently absorb every comment, because feedback can be incomplete, contradictory, or wrong.

The improvement loop

Think of the agent as working from a maintained playbook:

  1. The base instructions hold stable procedure and domain guidance.
  2. A reviewer leaves specific feedback on an output, including why the result missed the mark.
  3. An improver looks for recurring, well-supported patterns across feedback.
  4. It proposes the smallest instruction change likely to address the pattern.
  5. A person reviews the change before it becomes part of future runs.

This separates execution from learning. The working agent stays focused on the current task, while the improvement process can compare evidence across many runs.

Why small changes matter

Large prompt rewrites make it difficult to tell which change helped. Focused edits are easier to review, test, reverse, and explain. A good proposal should name the repeated failure, show the feedback that supports it, identify the instruction being changed, and predict the observable improvement.

For example, if a reporting agent repeatedly confuses a timing difference with an unexplained exception, a reviewer can explain the applicable business principle. The improver should propose a narrow addition to the reusable procedure rather than paste the entire review conversation into the next prompt.

Treat feedback as evidence

A thumbs-down signal indicates dissatisfaction but rarely explains the cause. Detailed expert feedback is more useful because it distinguishes a faulty fact, an unclear instruction, a missing source, and a disagreement in judgment.

Not every comment should become a permanent rule. Look for repeated patterns, check whether the feedback generalizes, and avoid building a brittle list of one-off prohibitions. Principles with a clear reason often transfer better than exhaustive exceptions.

Measure outcomes

An instruction changed is not proof that the agent improved. Track outcomes such as correction rate, reviewer effort, task success, or error recurrence. Compare performance on representative cases before adopting the update widely.

Practical checklist

  • Capture feedback where review already happens.
  • Ask what was wrong and why.
  • Group repeated patterns before proposing a change.
  • Prefer the smallest useful edit.
  • Keep proposals versioned and reviewable.
  • Test against representative cases.
  • Require human approval before permanent adoption.

Try it

Choose one recurring correction in an AI workflow. Write the underlying principle, a minimal instruction change, the cases used to test it, and the outcome that would demonstrate improvement.