Introducing the Traces Agent

Jun 15, 2026

We just launched the Traces Agent, a new way to search your team's session history in plain language. Ask it questions like "What did we get done this week?", "Should we skillify things we're all doing?", or "Who's using agents in the most interesting ways?" The Traces Agent reads through your trace history and answers directly.

How it works

The Traces Agent answers questions about your organization's AI-conversation history by looking things up directly in your traces. It never guesses from memory or makes answers up. INstead of relying on what a language model happens to "know," it runs a small set of precise lookups against your real trace data and builds its answer from what it finds.

It has four tools that mirror the agent.traces.com endpoints:

list. Recent trace metadata by agent, project, repo, status, and time, with optional AI analysis (summary, status, steps, badges, keywords) for triage. It never returns event content.
search. Grep across trace metadata and flattened event text (regex or literal).
show. Read bounded event content from a single trace.
lookup. Resolve a person, org/namespace, or registered agent to an id before filtering

It works from narrow scope to evidence: scope with list/lookup, discover candidate traces, read the actual events when you ask "why" or "which PR," then synthesize a brief answer. Every answer cites the traces it used as links, so you can click straight through to the source session.

It only ever sees what you can see

The agent's are scoped server-side to the caller's own access. This means you can only see public traces plus traces in namespaces you belong to. Private date from org you're not in is never reachable. Each chat is further pinned to a single owner namespace, intersected with your current membership, so a personal chat can't reach an org's private data and an org chat can't broaden to sibling orgs.

Available now for our most active users

Access it gated by activity: the agent chat unlocks for a namespace once it reaches a trace-count of 10.

How it works

It has four tools that mirror the agent.traces.com endpoints:

list. Recent trace metadata by agent, project, repo, status, and time, with optional AI analysis (summary, status, steps, badges, keywords) for triage. It never returns event content.

search. Grep across trace metadata and flattened event text (regex or literal).

show. Read bounded event content from a single trace.

lookup. Resolve a person, org/namespace, or registered agent to an id before filtering

It only ever sees what you can see