Git for the intent behind agent-written code. Records what your coding agent decided and why, and links every commit to it.
Why cogit
AI made writing code faster and moved the bottleneck to review. In a Harvard job-market working paper (Chen and Stratton, 4 August 2026; not peer-reviewed) covering 718 firms and 725,938 workers, AI coding agents wrote 30% more lines of code, made 20% more commits and opened 23% more pull requests, while resolved Jira issues and epics showed no statistically significant gain. The paper rules out an increase larger than 12% of the baseline mean. The review absorbed the difference: average review time rose 49%, comments per pull request 35%, the share of pull requests with changes requested nearly doubled (0.13 to 0.25), and the share of workers doing reviews rose 14%. Pull requests did not get larger, so size does not explain it.
A reviewer reading a diff sees what changed, not why: the goal, the decisions, and the options that were tried and rejected. That intent stays in the conversation with the coding agent, where the reviewer cannot see it. The author forgets it too: three weeks later, nobody remembers why the code went one way and not the other, and the agent starts every session knowing nothing about the last one.
At least one team already reviews this way: Anthropic's source control team still reviews every change to its system, but says it now examines "the intent of the change as opposed to reading every line of code."
cogit keeps that intent in the record and hands it over:
- the
Thought:trailer links each code commit to the message behind it, andcogit show <sha>opens that message cogit why <file>lists what earlier sessions decided and rejected for a filecogit prturns a branch into a review-ready description- the agent gets the same record at session start and before it edits a file
It does not promise a faster review; nothing has measured that yet. It gives the reviewer the intent they would otherwise have to ask for in comments, and gives the next session the context the last one ended with.
Sources
- Chen & Stratton, Artificial Intelligence in the Firm: Bottlenecks in Software Production (working paper, 4 August 2026)
- InfoQ's report
- Anthropic, Redesigning source control at Anthropic (13 October 2026)