Chain of Thought / Lab preview
One task.
Two interfaces.
Can an agent use the same evidence and rules through buttons or structured browser tools?
This experiment reuses the Would you merge this? scenario. Everything here is fictional. No code is merged or deployed.
The two modes expose the same fictional review task. In buttons-only mode, inspect evidence and make the requested decision through the page controls. In tools mode, a supporting browser can register structured tools against the same session. The result shows the evidence, decision, cost and route so you can compare what happened.
Try each mode from a fresh session and record whether the required evidence was inspected before the decision. A successful run of this fixed scenario is one observation; it does not establish a speed or reliability improvement for other tasks. The buttons remain available when tool registration fails.
The original review game provides the scenario. For the engineering discussion behind the exercise, read who reviews AI-generated code and explore the AI coding and evaluation hubs. Audit trail explains why the record of actions matters when checking a run.
Checking browser support. The buttons work independently.
The task
The agent fixed an export timeout for v3 clients and changed a failing assertion. Older v2 clients still depend on this service. Inspect the old-client failure and the rollback, then hold the patch for review. Report the result and its cost.
This fixed task tests interface reliability, not whether an agent independently chooses the best engineering decision.