Sessions

A session is one conversation with an agent or one run of an assignment, and the complete record of it. Every thought, tool call, and decision streams into the session as the agent works and stays there afterward, so what you read is the work as it happened.

That record is what makes delegation auditable. When a number in a report looks off, open the session and read what the agent actually did: which integration it queried, what came back, what it decided along the way. Six months later the session is still there, and it answers the question the same way.

A run, step by step

Take the Monday-morning pipeline report from the weekly report guide. At 6:00 AM the schedule fires, and a new session appears among the agent’s sessions with a live indicator. Open it and the work streams in: the agent reasons about the job, queries the CRM integration, works through the data, and produces the report — each step visible as it happens.

Partway through, you notice it is counting deals that closed last week. You type a correction into the composer without stopping the run, and the agent folds it into the work in progress. The finished report lands as a work product whose card links back to this session — the provenance of the artifact is the transcript that produced it.

Reviewing and steering

While a session runs you can stop it, send a message it reads after its current step, or switch the model tier and effort. Any passage of the transcript can take a comment, or a question to the agent in a side conversation. Checking your agent’s work covers each of these.

The microphone button in the composer dictates a message instead of typing it. Dictation uses your browser’s own speech recognition, which may send the audio to the browser’s maker (Google, in Chrome); Workprentice receives only the text. The button appears only in browsers that offer dictation.

Working with sessions

Sessions live in the sidebar; rename, archive, or delete them from the session menu, and search across them with ⌘K — handy when you remember the topic (“the churn analysis”) but not which agent or which week it belonged to. The sidebar filter narrows the list by agent, by who can access a session, or to sessions with failed runs.

Agents also hold sessions you did not start: each agent that answers in a channel or a direct message works in a background session of its own. The channel is the front door for that work, so an idle background session is hidden from the sidebar; tick Show background sessions in the filter to list them, and a live one appears either way.

If a conversation turns out to be a recurring need, Save as Assignment drafts an assignment from it — name and instructions distilled from what the agent actually did — so the session that got the work right becomes the template for repeating it.

Private and shared

Sessions are private to you until you share them. Share opens the sharing panel, where you can give the workspace access or add specific people. Anyone you share with can read the session, but only its owner can continue it: a teammate who wants to build on it forks it — from the latest message, or from any earlier point — into a private copy of their own, while your record stays exactly as you left it. Forks that their owners share appear as markers on the original’s timeline at the point they branched.

Multi-agent work

Agents can delegate: a session may hand work to subagents, which appear inline as cards showing each subagent’s task, its live activity, and its result, with a link that opens its full transcript. The parent session stays readable because the detail lives one level down: you can see that research was delegated and how it went, and open the subagent only when you want the step-by-step. Mention another agent by typing @ in the composer to bring it into the conversation directly. Delegation is bounded by the collaboration guardrails, so an agent’s reach stays within the limits you’ve set.

Wakeups

An agent can schedule its own follow-up — “I’ll check the deploy again in an hour.” A pending wakeup shows as a pill on the session with the time, and you can cancel it. Wakeups suit the agent’s own short-horizon follow-through: checking whether something it kicked off has finished, looking again once new data should exist. For work you want repeated, use an assignment instead — assignments carry a schedule, pause controls, and a per-run audit trail, where a wakeup is a single deferred next step.