New Adapter: Prime Agent

Aug 11, 2026

Traces now supports Prime Agent, the self-improving RLM harness from Prime Intellect. Your Prime Agent sessions are now captured in Traces

What is Prime Agent?

Prime Agent is a self-improving RLM (Recursive Language Model) harness built for long-running, autonomous coding tasks. It executes through IPython and maintains a self-modifiable harness state, so it can refine its own behavior mid-run. Prime Agent also spawns nested RLM sub-sessions to break work apart. Traces can capture all of it.

How Traces reads Prime Agent

Prime Agent writes each root session as a JSONL event log under ~/.prime/agent/sessions/, and child transcripts under /.prime/agent/session-artifacts/.

Each JSONL line is discriminated by type. Traces normalizes into a commmon structure:

Full turns. User prompts, assistant thinking blocks, assistant text, tool calls, and tool results all map to their own trace events, capturing the complete back-and-forth.
IPython execution Tool calls and results carry the tool name (e.g. ipython) and mark a result as failed when Prime reports details.status: error or isError.
Self-modifying harness state. Prime Agent emits prime-agent.refinement records when it edits its own harness, which captures the refinement's scope, summary, rationale, harnessStatePath, and the exact before/after appliedEdits. Traces surfaces these as agent_state.update system events so you can see the agent improving itself.
Nested RLM subagents. Child sessions record their parentSession and rlmDepth. Traces walks those links so a session and every sub-session it spawns stay connected.

Session titles come from the first user prompt, git metadata (remote URL, branch, commit) is picke dup from the session header when present.

What you can do now

With this update, you can:

Share Prime Agent traces. Traces auto-detects Prime Agent from ~/.prime/agent/sessions/ and publishes the full execution with traces share.
Resume sessions. Traces hands the session ID back to Prime Agent via prime-agent --resume <id>.
Watch live. Traces polls the active session and streams new events as they land, so a running autonomous taks shows up in real time.

Get started

Update your CLI.
Run traces share to publish a Prime Agent session and get a link.

What is Prime Agent?

How Traces reads Prime Agent

Prime Agent writes each root session as a JSONL event log under ~/.prime/agent/sessions/, and child transcripts under /.prime/agent/session-artifacts/.

Each JSONL line is discriminated by type. Traces normalizes into a commmon structure:

Full turns. User prompts, assistant thinking blocks, assistant text, tool calls, and tool results all map to their own trace events, capturing the complete back-and-forth.

IPython execution Tool calls and results carry the tool name (e.g. ipython) and mark a result as failed when Prime reports details.status: error or isError.

Self-modifying harness state. Prime Agent emits prime-agent.refinement records when it edits its own harness, which captures the refinement's scope, summary, rationale, harnessStatePath, and the exact before/after appliedEdits. Traces surfaces these as agent_state.update system events so you can see the agent improving itself.

Nested RLM subagents. Child sessions record their parentSession and rlmDepth. Traces walks those links so a session and every sub-session it spawns stay connected.

Session titles come from the first user prompt, git metadata (remote URL, branch, commit) is picke dup from the session header when present.

What you can do now

With this update, you can:

Share Prime Agent traces. Traces auto-detects Prime Agent from ~/.prime/agent/sessions/ and publishes the full execution with traces share.

Resume sessions. Traces hands the session ID back to Prime Agent via prime-agent --resume <id>.

Watch live. Traces polls the active session and streams new events as they land, so a running autonomous taks shows up in real time.