When you build with agents, every change comes out of a detailed conversation about how it came to be. Those conversations reference the code, the PRs, and each other. Polygraph records them, so humans and agents in your organization can see what is going on.
Every session emits a trace: the visible record of what the agent did and why. Where memory is the durable state a session preserves, the trace is the observability layer on top of it: prompts, decisions, tool calls, and files touched, in order.
When you'd open it
Section titled “When you'd open it”- Live: check what an agent is doing right now, mid-run, including the child agents working in auxiliary repos.
- Replay: walk back through what the agent did and why, after the fact.
- Hand-off: share the trace with a teammate so they pick up the reasoning, not just the diff.
- Reference: a new session reads a prior session's trace to inherit context.
What gets captured
Section titled “What gets captured”- Every prompt and agent response.
- Every tool call (file read, command run, MCP call) with arguments and result.
- Every git operation (branches, commits, pushes, PRs).
- Linked references (issues, PRs, sibling sessions).
Viewing the trace
Section titled “Viewing the trace”Each session has a shareable link in the Polygraph web UI that renders the trace timeline. Use polygraph session show --details to dump the relevant fragments to the terminal.