What AI agent workflows actually survive 6 months on a real team
A development team shared which s have stuck around for six months of real use, versus ones that quietly died. The clearest survivor is : coderabbit reviews every before a human does, catching small mistakes so the team's senior engineer only needs to check design decisions instead of nitpicking. That engineer used to be the bottleneck; now they aren't.
Second, after a merge, an agent drafts test cases and a team member keeps the good ones — test coverage rose without anyone being formally assigned to own testing. Claude Code handled a REST-to-GraphQL migration by mechanically converting roughly 40 straightforward endpoints, while humans redesigned the five tricky ones by hand. A weekly dependency audit that automatically opens upgrade also stuck, covering a job nobody used to do.
The weakest link is syncing: an agent drafts doc updates based on merged , but a human still has to read and approve them, and some weeks nobody does — pushing the stale-docs problem up one level instead of solving it. Every workflow that died shared one trait: letting the agent both do the work and decide it was finished. Auto-merging on passing tests lasted only two weeks before being scrapped.
Key points
- coderabbit reviews every PR before humans, removing the senior engineer as bottleneck
- Agent drafts test cases post-merge; team keeps the good ones, raising coverage with no dedicated owner
- Claude Code mechanically converted about 40 REST endpoints to GraphQL while humans handled 5 tricky ones by hand
- Weekly automated dependency audit now opens upgrade PRs on its own
- Workflows where the agent decided its own completion, like auto-merge on green tests, died within two weeks