Auto Lab · DocsBack to app

Chat and runs

The goal's one conversation across the web and your channels, the task agents it hands work to, and the sessions they work in.

Chat is the goal's single running conversation. You and the goal's other admins talk to its agent here, from the web app or from any channel connected to the goal. Goal members can read Chat but not send messages. This page covers what you can do in Chat, what shows up in it, how the agent hands work to task agents, and where sessions fit.

One conversation per goal

Select Chat in the goal's rail. Messages from the web app, Slack, Microsoft Teams, iMessage and email all land in the same conversation, so the agent has the whole story wherever you write from. Its answer goes back to the place the message came from, and the web page shows everything. To connect a channel, see Channels.

A message from someone who is not on the goal, such as an email from outside your organization, is answered in Chat only, with limited tools. Nothing goes back to that sender until someone on the goal gives the go.

What you can do in Chat

ToDo this
Ask or hand over workType in the composer and send. Say what you need in your own words.
Steer while it worksSend another message. The agent reads it at its next step instead of waiting for the current turn to end.
Answer a specific messageHover the message and select the reply arrow, double-click it, or long-press it on a phone. The composer names the message you are answering; press Esc to cancel.
Attach filesSelect the paperclip. Up to 10 files per message, 20 MB each: photos, PDFs, text and data files, Office documents. A message can be files alone.
Stop the agentSend /stop. It ends the running turn. Task agents that are already running carry on; cancel them on the Runs page.
Shorten the historySend /compact. The agent summarizes the older part of the conversation and keeps the recent part word for word.

The agent looks at photos directly. To read a document, it hands the file to a task agent.

/compact changes what the agent carries into each turn, not what you see: every message stays on the page, and the agent can still search the full history. It also compacts on its own when the conversation grows long. Either way, Chat shows History compacted.

When the agent answers something other than your last message, such as a task reporting back or a reminder firing, its reply quotes the message it belongs to. Select the quote to jump to the original.

What shows up in Chat

Chat also shows what happens between messages.

You seeWhat it means
Needs your approval, with Approve and DenyThe agent wants to take an action your rules gate. The card names the app and action, its risk (read, write or destructive) and the exact values. See Approvals and autonomy.
A question above the composerThe agent needs a decision or a fact. The card stays pinned until you answer, so it never scrolls away. Pick an option, type your own, or select Skip to let it continue without an answer.
… task cards, such as Research taskA task agent is working. Progress lines fill in, then its report. Open run opens it on the Runs page.
Connect …The agent needs access to an app. Connect opens the connection page in a new tab. Once the app is connected, Chat carries on with what you asked.
Learned: … with UndoThe agent saved something it learned. Undo takes it back. See Learning.
Scheduled reminder, Scheduled check-up, Scheduled routineOne of the goal's schedules fired. The messages below it are that run. See Plans, schedules and triggers.
Reply arrivedSomething the agent was waiting for came in, such as an answer to an email it sent.
Delivery to … failedA reply could not reach Slack, Teams, iMessage or email. The agent does not claim a message was delivered when it was not.

Older messages load a page at a time: select Load earlier messages at the top.

Task agents and runs

To keep the conversation focused, the agent hands pieces of work to task agents. Each hand-off is a run. A run shows as a card in Chat that fills in as the task agent works.

Task agentWhat it doesLimit
ResearchAnswers one question. It reads the goal's connected apps (read only), its memory and, when it needs them, its files, and searches and reads the web. It changes nothing in your apps.60 steps, 25 minutes
DocsDrafts or updates one document or message. It can read and write in connected apps; a write your rules gate comes back to Chat as an approval card.40 steps, 20 minutes
Mail readerTurns one email into a short summary: what it says, what it asks, the dates. It has no tools, so it cannot act on what the email says.1 step, 1 minute
BrowserUses one website in a browser of its own, signed in with a saved login when there is one. It never sees your password. A purchase stops at the final review and waits for you to confirm.60 steps, 20 minutes
CoderOpens a session on its own branch of the goal's repository, makes the change and ends with a change request for you to review. The card links Review change request #….60 minutes

A few rules apply to every task agent:

  • It cannot ask you anything. When it needs you, it stops with Needs input and the agent asks you in Chat.
  • It cannot set schedules or start other task agents.
  • A goal runs up to 10 task agents at a time, counting queued ones. At the limit, the agent waits for one to finish or cancels one.
  • If the agent is waiting on your approval or answer when a run ends, the run's report waits until you decide.

The Runs page

Open Runs from a task card's Open run, or from the ⋯ menu on the sidebar's recent list while Chat is selected. Each row shows the role, the brief, the status, when it started, how long it took, the steps it used out of its limit, tokens, cost, and where it ran:

EnvironmentMeaning
SharedIt ran without a machine of its own.
Own boxIt had its own machine, for files or a browser.
WorkbenchIt worked in a session. Coder runs work this way.

Select Cancel to stop a queued or running run. Select a row to open its brief, its report and What the agent did, step by step.

Watch the browser live

While a Browser run is working or waiting for input, open it to see a Live view: the page as the agent sees it. If the agent is stuck on a check it cannot clear, you can click the picture, Type a short text such as a code, or Press one key.

Sessions: the workbench

Chat is the conversation; sessions are the workbench. A session is an isolated cloud machine with its own branch of the goal's repository. On a goal that runs in Chat you do not talk in sessions. The agent uses them for hands-on work:

  • a Coder task works in a session and ends with a change request
  • a task agent that needs files or a browser gets a machine of its own
  • a scheduled trigger from Brain › Triggers starts a session for each run

To see them all, open the ⋯ menu on the sidebar's recent list and select All sessions. Every row carries a Workbench tag, and a session a task agent opened links back to its run. Filter by Mine or All sources, group by status or source, or search by title. Manage lets you rename, share, restart, stop or delete sessions. New session opens one directly, for example to work on files with an agent; that conversation stays in the session and does not appear in Chat.

A session's machine stops when it has been idle for a while and keeps its files. Only work committed to the session's branch lasts after the session is deleted. Each session has its own access setting, Session access: Only you, Specific people, or everyone on the goal. The person who started the session changes it. For sessions a trigger starts, set it on the trigger.

Bring an earlier session into Chat

A goal that worked in sessions before it moved to Chat opens with an empty conversation. Select Import the last session as history to copy the text of the most recent finished session into Chat. The agent can then refer to it and search it. Nothing runs, tool details are left out, and each session imports only once. If the agent is busy, try again when it finishes.

Older goals without Chat

A goal that has not moved to Chat opens Chat as a new session. Each conversation there is its own session, listed in the sidebar, and the agent works in that session's machine. These goals have no Runs page and no Brain › Schedules; recurring work runs from Brain › Triggers.