Audit trails for AI coding agents (Claude Code, Codex, Cursor)
Updated · 8 min read
What tools give me an audit trail for AI coding agents like Claude Code and Codex?
Start with what you already have: git history, pull requests and CI show what changed in the code. Each coding agent adds its own logs (Claude Code exports OpenTelemetry metrics and events, Codex keeps session logs, Cursor offers admin audit logs on business plans). What none of them give you is one record across tools of why the work was done, who approved it, and what evidence showed it worked. That is the part OrgX keeps.
What an audit trail for coding agents should answer
- What was the agent asked to do, and by whom?
- What did it change (files, commands, pull requests, deploys)?
- What was it allowed to do, and did it ask before anything risky?
- How was the result checked: tests, review, a person's sign-off?
- What is still open, and what should the next agent or person know?
What each layer records
- Git and your code host: every commit, pull request, review and merge. This is the most trustworthy record of what changed, and it already exists.
- CI: which checks ran on which commit and whether they passed.
- Claude Code: permission prompts and hooks on the machine, and OpenTelemetry export of usage metrics and events that you can send to your own monitoring. Teams on Anthropic's business plans also get usage analytics.
- Codex: an approval policy and sandbox per session, and local session logs of the conversation and the commands it ran.
- Cursor: agent chat history in the editor, command allowlists, and admin audit logs on business plans.
- Observability tools (LangSmith, Langfuse, Arize Phoenix): traces of model and tool calls. Strong for debugging agents you build; they do not know which business task a coding session served.
The gap: one record across tools
Most teams use more than one agent. A feature might start in ChatGPT, get built in Claude Code, get fixed in Cursor and get reviewed by a person. Each tool keeps its own log in its own place and format, and none of them records the decision behind the work or who accepted it. When a client or a teammate asks "who checked this?", the answer lives in someone's memory.
How OrgX fills it
- Agents connect through the OrgX MCP server and plugins, read the task and the decisions already made, and write back what they did.
- Each finished piece of work carries a receipt: what was asked, what authority the agent had, what changed, the evidence (pull request, test run, URL), and whether it was verified.
- The receipt format is open: Agent Work Receipt v0.1 is a JSON Schema you can emit and validate without an OrgX account.
- Sends, deploys and spend stop for a person, and the approval sits in the same record as the work.
- Honest limit: capture from Claude Code today needs the OrgX hooks installed alongside the plugin, and receipt coverage from Claude Code sessions is lower than from Codex. Git stays the source of truth for code changes.
Which record answers which question
| Question | Where the answer lives | Across tools? |
|---|---|---|
| What changed in the code? | Git, pull requests | Yes, per repository |
| Did the checks pass? | CI | Yes, per repository |
| What did the agent run on my machine? | Claude Code telemetry, Codex session logs, Cursor history | No, one log per tool |
| Why was it done, and who approved it? | OrgX record and receipts | Yes |
Questions people ask
Do I need a separate tool if I already have git history?
For small solo projects, git plus pull request descriptions may be enough. You need more when several agents and people share the work, when risky actions need approval, or when you must show a client or auditor who checked the work.
Does OrgX record every command an agent runs?
No. OrgX records the work: the task, decisions, approvals, evidence and outcome. Command-level logs stay in each tool (Claude Code telemetry, Codex session logs). Link to them from the receipt when they matter.
Can I use the receipt format without OrgX?
Yes. Agent Work Receipt v0.1 is a published JSON Schema with example receipts for Codex, Claude Code and OpenClaw, and a validator that works without an account.
Related
Proof for AI-delivered work
OrgX is proof for AI-delivered work — the operating record your agents and tools share.