History & Replays

History and replays

Every run and Inspect session is kept whole. History lists them all; each one can be replayed, read call by call, and shared as a link or an embed.

Note

History is for signed-in accounts, and lists the runs and Inspect sessions of the workspace you have open. Sign in to keep yours.

History

History lists every run and Inspect session in the workspace, newest first, 30 to a page. The navigation also shows your last few runs at the store in scope; History has the rest. Each row shows:

  • Kind: a run or an Inspect session
  • Task: what the agent was asked, or the tool you called
  • Store, and for a run, the model
  • Calls, seconds and cost
  • How it ended: Searched only, Built a cart, Reached checkout, Tried to place the order, Order placed, or Failed (see Understanding results)
  • When it ran

Click a row to open the run or session.

Filtering

  • What to show: Everything, Runs or Inspect sessions
  • Store: one store, or Every store. The store you are working on is marked "(in use)".

Replaying a run

Open a finished run and the bar under the header says Recorded. Press Replay to play the run back at its real pace, or at 4×; Step jumps to the next event, and the slider moves to any moment. The chat, the Trace and the Timeline show only what had happened by then. A replay is read-only: you are watching a finished run, not running it again.

While a run waits for the buyer (Your turn), the bar has a Replay button too.

Everything the run used is kept with it, not read again today. The Discovery and negotiation tab shows what the store's UCP profile said and what the two agreed at the time; the Info tab shows the task, the setup, the buyer details, the tools and how it ended. Use the Trace to check that a store's answers are well formed, and the chat to see why the agent made a choice. Runs recorded before negotiation was kept say so instead of showing empty blocks.

A run through the store's website also keeps what the browser saw: The agent's browser pane shows the page the agent opened, a screenshot after each step that changed the page, and the checkout where it stopped. See WebMCP: the store page.

Sharing a run

Click Share in the run's header, then make a public link. Anyone with the link can see the run, without signing in. The dialog shows the card people see when the link is shared, the Link with Copy link, and the Embed code with Copy embed code. Once a run is shared, the header button reads Shared.

  • Shown: the buyer chat as the buyer saw it, every call the agent made and the store's answers, and the model, timings and tokens per call.
  • Never shown: the system prompt, the cost and who ran it, and the buyer's details and checkout links.

A share is a copy of the run as it is when you share it. Later turns don't change it; stop sharing and share again to update it. The same copy is available as JSON at /s/{id}.json. Stop sharing turns off the link and every embed of it; the run itself stays in your History. A website run's screenshots appear on the shared page too.

Sharing is useful for:

  • sending a run to a colleague to debug a store issue
  • sharing results in a Slack thread or a GitHub issue
  • including evidence in a UCP compliance report

Embedding a run

For documentation, blog posts or presentations, paste the embed code from the Share dialog into your page. It is an iframe of the shared run: viewers can scroll through the chat and open the calls without leaving your page.

Tip

Combine sharing with the Funnel & Outcomes reading for a before-and-after. Run a task, fix what the run shows, run it again, and share both runs to show the difference.

Something wrong on this page?