Checking Your Agent's Work

Every piece of agent work leaves a record you can check: the deliverable links to the session that made it, and the session shows each step the agent took. This page covers reviewing a result, asking about a step, correcting the agent so the fix sticks, stepping in while it works, and the two safety nets that work while you aren’t watching: the Inbox and Guardian.

From the deliverable to the steps behind it

Start from the thing you care about. Documents, spreadsheets, charts, and other files your agents produce collect on the Work Products page, and each card has View session, which opens the conversation or run that produced it. A document that came from a session has the same View session link in its header.

In the session, the agent’s work sits above each reply as a collapsed line: a short summary (or Worked for and a duration) followed by a count of thoughts and tool calls. Click it to expand the steps: each file the agent opened, each query and calculation it ran, each decision along the way. Say the pipeline report shows a total you don’t believe. Expand the work and look for the step where the agent read the CRM export or queried the CRM; you can see what it asked for, what came back, and what it did with the result, so you can tell whether the data was wrong or the agent misread it.

Asking about a step

Select any passage of a transcript and two actions appear: Comment and Ask agent. A comment leaves a note anchored to that passage, which you can later resolve or reopen, so a teammate can find the step where a number went wrong.

Ask agent opens a side conversation with the same agent about the selected passage, docked beside the transcript. Ask why it chose that source, or what would change if the date range were different; the main conversation is untouched while you do. When the side conversation reaches something the main one should act on, Commit to conversation drafts a summary from the agent’s last reply, which you edit and commit as your next turn. Nothing crosses over until you commit, and the agent cannot commit on its own.

Correct it once and the correction sticks

When something is wrong, tell the agent in the session instead of fixing the output yourself: “our fiscal year starts in February,” or “leave out closed deals.” The agent redoes the work and saves the lesson to its memory, which shows as a Remembered step in the transcript. Every later session starts with it, so you don’t repeat the correction next week. If no Remembered step appears, ask it to remember the rule.

Memory is visible on the Memory card of the agent’s page, where each entry keeps its history and can be restored if an update went wrong. See Memory and storage.

Stepping in while it runs

A running session shows a live indicator and the agent’s steps as they happen, and you can intervene at any point:

The stop button in the composer (Stop generation) halts the agent.
Send a message while it works and it is queued; the agent reads it after its current step and takes it on board without stopping. For a small correction this beats stopping and starting over, because the agent keeps the work it has done.
The model and effort picker in the composer changes the tier and reasoning effort from the next turn on, which is how you hand a hard step to Strategic without changing the agent’s defaults.

A session that ended unexpectedly is marked and offers Resume session, so the work picks up where it stopped. A session budget, unset by default, pauses a run for your review once it has spent that much, and the session offers to resume with a larger budget or none; see How agents run.

The Inbox: when the agent needs you

Most runs happen when nobody is watching. When an agent needs an answer, is blocked, or has something worth showing, it sends a message to your Inbox with a priority instead of guessing or failing silently. Questions it needs answered usually arrive as Urgent. A typical one: “The export has two revenue columns, booked and recognized. Which should the pipeline report use?”

Reply in the thread and the waiting session continues from your answer. The Inbox can also reach you in Slack, by email, or by text, and a reply there works the same way.

Guardian: a check before anything leaves

Guardian, on by default, reviews what an agent is about to push or post for confidential information and holds the call with a reason when it objects; see Guardian. It is one of an agent’s guardrails, which also decide which integrations and shell commands an agent may use, whether it may hand work to other agents, and what it may spend.