Finding a patch is one part of fixing a bug. Engineers also need to know whether it covers the reported failure, works with existing behaviour and will be easy to maintain.
We compared saved investigations of an Apollo Kotlin cache bug: Codex working alone and Codex working with Coral Code. Both produced a working fix. The Coral-assisted submission added checks for more ways an application could use the cache.
Test the behaviour a user depends on
Applications could receive a correct network response, save it and then fail to read it back from the cache. A useful test had to exercise that full path, rather than stop at parsing the response.
Both investigations reproduced the defect and tested their fixes. Codex alone passed four focused regression cases. Codex with Coral passed five, covering different cache-key strategies and reads after the cache already contained data.
Use other agents to question the work
Coral’s agents contributed findings during the investigation and raised a possible side effect. The coding agent checked that concern against the source and tests, then explained why it did not apply.
Your reviewers get a diagnosis, a patch and evidence from checks of surrounding behaviour.
Keep the strongest parts of each solution
The standalone Codex submission tested a useful change in fragment order. Coral’s submission covered more cache states. Apollo’s maintainers merged a smaller implementation that also passed both saved regression suites in our replay.
Coral-assisted investigation broadened the checks around this fix. This comparison covers saved code and test results from runs with different starting states.
Inspect the engineering evidence
Read the reported bug and maintainers’ merged fix. Explore Coral Code for the developer workflow and the project documentation for setup.