The Git Forge for Agent-Human Collaboration

Discussion overview

The thread focused far less on a new Git forge than on its underlying wager: when coding agents generate most changes, should teams review the agent’s trail of reasoning rather than a pull request diff? Several commenters grant that agent-produced diffs are cheap while their rationale evaporates quickly; preserving sessions with commits could be genuinely useful weeks later during debugging, incident response, or archaeology. The proposed graph, automated evaluations, and continuous knowledge synthesis also struck some readers as a plausible direction for AI-assisted review.

But confidence is low that this warrants replacing ordinary review. Skeptics argue that a PR is already the compressed, inspectable result of work, and replaying a long agent session resembles asking reviewers to watch hours of a developer’s screen recording. That creates more noise, not better judgment, especially for teams unwilling to trust AI-generated code without reading it closely. Proprietary-context exposure is another hard blocker: organizations need a compelling reason to hand over agent prompts, transcripts, and internal reasoning when they can store them themselves. Readers also pushed back on the startup’s sweeping claim that humans will soon write little code, seeing it as marketing rather than a settled technical future. The product has enthusiastic early users, but the conversation remains cautiously skeptical about the workflow shift and its trust model.

Points of disagreement

The core dispute is whether agent session history should displace the pull request as the main review artifact. Supporters say agent-written code loses its why almost immediately, so commit-linked context, checks, and evaluations preserve the evidence needed to understand failures later. Critics say a diff is a deliberate compression of the work; reviewing sprawling agent transcripts is an expensive, high-trust ritual that few engineers will actually perform.