Agent sessions: see how the coding agent got there

A step that hands work to a coding agent used to leave behind only the agent's last word. The sandbox it worked in, and the agent's own history, went when the run finished. Now Murmurator keeps the session.

What's new

• Agent sessions. Every workspace.agent call is recorded, however it ended: the prompt, what the agent said, each command and file edit with its output and exit code, and every model call's tokens. Secrets are scrubbed out, as they are from step logs.
• On the run page and on their own. A step links to the sessions it ran, and the new Agent sessions page lists them across runs, filtered by workflow, step and outcome. Failed commands open by themselves.
• The assistant reads them. Ask it to improve an agent step and it starts from where the agent struggled. Fixes it suggests for failing evals use them too.
• Over MCP. list_agent_sessions and get_agent_session give your own tools the same record.
• Longer agent answers. A long codex session used to come back cut short at 20 KB, and with it the agent's final answer. The whole stream is read now.

Read more in Workspaces.