What “fix this bug” left out
A system can complete the steps it was given while missing the responsibility the business thought it had handed over.

At Duolingo, some bugs begin with a shake. Admin and beta users can shake a phone to open a report form and send screenshots, logs, device details, and other debugging information. Reports submitted by admin users automatically become Jira tickets. The reporting step was quick. With more tickets arriving than engineers could resolve, the backlog kept growing.1
That made the company's experiment with an AI bug-fixing agent sound almost inevitable. The request at the center of the project was only four words: “Claude, fix this bug.”1
The request was short because it left out most of the work.
An engineer receiving a bug report has to reconstruct what happened, understand how the feature is supposed to behave, find the relevant code, test a theory, make a change, and decide whether the result is safe to move forward. Those steps do not fit inside the four-word request.
Duolingo built a process around the request. It supplied the ticket and relevant repository, then had the agent gather product context and research the feature before changing code. That research step came after convention-breaking changes created extra work for reviewers. The team also added self-review, had the agent monitor automated checks and revise its work when they failed, and automatically created beta builds that anyone could install to verify a fix.1
Microsoft found the other half of this problem while building Copilot Cowork. One formatting check had been handed an empty list of files, so it inspected nothing and passed. The team responded with two rules: every check had to report how many files it examined, and every new check had to include a test that deliberately made it fail.2
Read together, the two companies’ engineering accounts make a practical point about delegation: a system can complete the steps it was given while missing the responsibility the business thought it had handed over.
An invoice makes that easier to see. An AI system might extract the supplier, amount and due date correctly, then find that the amount does not match the purchase order. The business responsibility is no longer data entry. Someone has to decide whether the mismatch is acceptable, whether the invoice should be held and who has authority to approve the exception. The system can collect the evidence and route the case. It should not quietly invent the authority to settle it.
Auxeon’s method addresses this before the system is built, by identifying the records it may rely on, the situations that must stop for review and the person who can decide what happens next. The evidence travels with the exception to that person, and the decision is recorded.78 The aim is to keep investigation and accountability inside the work instead of passing them to whoever discovers the gap later.