The industry has spent the last few years optimizing AI agents’ code-generation capabilities. The focus has been on expanding context windows, fine-tuning models on repository-specific data, and developing complex prompting strategies. This has undoubtedly produced more capable coding agents. However, for most teams, that code-generation capability has not translated into significant gains in productivity.
Most engineering teams are stuck in a manual workflow. The agent generates the code, tests it locally, and submits a PR to the developer…
