An AI coding agent gets a ticket that says, “Cancel a subscription when the customer requests it.” The agent changes the API, updates the interface, and writes a test. The pull request passes CI. Then a reviewer spots a decision nobody made. The code removes access immediately, including access the customer has already paid for.
This is a useful failure to test when introducing AI coding agents. A vague requirement can produce a consistent plan, implementation, and test suite. The team needs a way to catch the assumption before it travels through all three.
What should the agent hand off for review?
Before coding, surface the open decisions. For the cancellation ticket, the plan should state when access ends, what happens to the next renewal, and how the system handles a cancellation submitted while renewal is already processing. An engineer can resolve those questions with the product owner before the agent edits the application.
**During testing, check the approved rule independently. **Suppose the agreed behavior is that cancellation prevents renewal and preserves access through the paid period. A reviewer can add a test for a customer cancelling halfway through that period. The test should check access as well as the subscription status. A component test that confirms the cancellation message is displayed might miss the real defect.
Before merge, make production ownership explicit. The pull request can link to the decision, show which checks exercise it, and name the engineer responsible after release. If support reports that customers are losing access early, the owner has a record of the intended behavior and where it was verified.
GeekyAnts documents this kind of handoff on selected projects. Agents draft requirements, split work into tasks, implement changes, and run tests. Engineers approve the plan before coding and decide how failed tests should be handled. The agent proposes its changes on a separate branch, and the engineer who merges them owns the change in production.
The companion video shows how those reviews fit into the workflow, from approving the plan to merging the agent’s changes. Watch the requirements-to-release walkthrough : https://www.youtube.com/watch?v=brUkhEsnyIU
How four other companies approach the workflow
Thoughtworks uses AI/works to combine requirements and existing system context into an evolving specification. Teams can review it before generating implementation tasks and tests.
EPAM reports using CodeMie to carry requirements and architecture context into implementation, testing, and release. Its feedback assistant can turn a raw request into a Jira-ready item for product review.
Globant documents a Test AI Agent that drafts and executes test plans. A team evaluating that approach can inspect whether the plan covers timing, retries, and exceptions alongside the standard customer journey.
Accenture worked with GitHub to study Copilot in enterprise development. The research combined a randomized trial, adoption analysis, and developer
A small change to the pull request template
Three fields can make an AI-assisted pull request easier to review:
- Approved behavior. Link to the requirement or decision.
- Evidence. Include a test result or reproducible check for the riskiest scenario.
- Production owner. Name the engineer responsible for verifying the change after release.
For the cancellation example, the strongest evidence would show that a customer who cancels midway through a paid period retains access and avoids a new charge. If a test expects access to end immediately, the reviewer can compare it with the approved requirement and decide whether the test or implementation needs to change. Recording that decision gives the next engineer working on billing a reason for the behavior.
Which approval point catches the most issues on an AI-assisted project?
Top comments (0)