A Digital Employee Is Not Done Until Completion Is Independently Accepted
Computer-use agents need a completion contract that separates process evidence, business outcome and failure classification. The worker may claim completion, but it must not accept its own claim.
Core judgment
Completion is a governed claim, not the last action, a final screenshot or a model-generated sentence.
Why this is not a point feature
A final screenshot can show an outcome artifact without proving required steps, side-effect limits or business authority. A process score can show execution quality without proving the requested business state. A failure label can assign responsibility without providing retry safety. These are separate claims and need separate evidence.
Minimum deployable architecture
Use a versioned Completion Claim containing the expected outcome, process evidence, deterministic state readback, optional learned-verifier advice, failure and side-effect classification, and an independent acceptance decision. Preserve disagreement instead of forcing a binary answer.
Boundaries and counter-evidence
The supporting research is web-task specific, uses a 246-trajectory benchmark and still reports an 8% external false-positive rate. It does not demonstrate rollback, compensation, transactionality or enterprise incident reduction.
Engineering conclusion
Adopt the separation pattern first: claimant, evidence contract, verifier and acceptor. Treat learned verification as advice alongside deterministic business checks and required human authority.
Visualization note
The diagram represents control boundaries and state relationships. It does not present experimental results or invent quantitative comparisons absent from the Research Object.
Evidence and citations
- Research Object: the sole analytical input, including judgments, uncertainty, counter-evidence and engineering implications.
- Reading Result: the evidence boundary and source-trace record behind the Research Object.