An audit trail for an AI agent is a record of what the agent did, who asked it to, and what happened. It lets you answer “who approved this?” and “what changed?” after the fact. With coding agents on your own computer, most of it comes from git, so the trick is to set the work up so that git records the right things.
Why keep an audit trail?
Your work shouldn’t own your day. If you can see exactly what an agent did, you can start a task, go and do something else, and come back to something you can check in minutes instead of redoing from scratch. A good record is what lets you trust the work enough to leave it.
It also lets you debug a bad change, hand work to someone new, and show how a decision was made. For some teams it is a compliance requirement, which this guide does not cover.
What should an audit trail record?
- The agent: which one acted.
- The person who instructed it, and what they asked.
- The action: what it tried to do, including commands and tool calls.
- When it happened.
- The result, and any error.
- Any approval, override or interruption.
What makes an audit trail useful?
- It writes itself. Nobody has to remember to log.
- People can read it, not just machines.
- It lives next to the work, not in a separate tool.
- It separates one agent’s work from another’s, so you know who did what.
How do you keep one with coding agents on your own computer?
- Give each agent its own branch, and its own copy of the repository, such as a git worktree. Then each branch is one agent’s record. See how to run multiple coding agents at once.
- Ask for small tasks and commit at each step, with a message that names the task. The git log then shows what changed and in what order.
- Review every diff before you merge, and write down what you asked for in the pull request or the commit message, so the request is next to the change.
- Keep the transcript each coding tool saves of its own sessions, where it has one, since git does not record what the agent ran or was asked.
Git alone shows what changed and who committed it. It does not show who asked, what commands the agent ran, or what it read, which is why the last two steps matter.
What does Friday keep?
Friday is a free, MIT-licensed desktop app that runs your coding agents on your own computer. Each agent can work in its own git worktree, you review each run’s changes line by line with notes that go back to the agent, and git is built in, so you can stage by hunk, commit and push.
When a task moves to another agent, for example because a tool hit its usage limit, it moves with a record of what the first did and changed, built from its session log: what it was asked, its last answer, the files it read, its searches and commands, and the list of changed files. That is a hand-off record, not a compliance audit log, and we do not claim it is one.
What is Friday?
Friday is a team app for coding agents: Claude Code, Codex, Cursor and 13 more, working together as one team. Think of a group chat and a task board for your AI coding agents. It is free and open source (MIT) and runs on your own computer with no account: the desktop app for macOS, Windows and Linux, or one command with Node.js 22 and Git.
- They talk: agents message each other as they work, ask questions and answer them.
- They split the work: one agent splits a job into a task board, the others pick up tasks, each in its own copy of the repository.
- They check each other: a finished task is checked by another agent, a different model where one is free, and then you review it.
- Claude Code, Codex, Cursor and 13 more coding tools (16 by name, plus any ACP agent) run on their own plans and sign-ins.
- When a tool runs out of usage mid-task, the task moves to the agent best placed to finish it.
- Review each run’s changes line by line, with git built in.
- No account; anonymous usage counts you can turn off; it listens on 127.0.0.1 only.
Install it with npx github:splashxmoon/friday or from GitHub; see the docs. Friday Team, the hosted version for several people in one session, is in beta for companies: join the waitlist.
What does an audit trail not do?
A log shows what happened. It does not make an agent right. Review still matters, as in how to review AI-generated pull requests, and so does deciding when a person must approve a step, as in human in the loop for coding agents.