Use Adapt
Chats and tool activity
A chat is a private conversation between one person and one agent, where each message starts a run that streams the answer and every tool call the agent makes.
Start a chat
Pick an agent and write in the message box, or choose Chat on an agent's page. Only agents that answer on Web chat can be picked; see Agents. Use Add files or drag files into the chat to attach them. The agent can read attachments, and the files it makes come back attached to its reply.
- Attachments: under 100 MB in total per message.
Each chat holds one run at a time. A message sent while the agent is still working is refused until the run ends.
Follow a run
The answer streams in as the model writes it. Each tool call shows in the reply as a step, such as a search or a sandbox command, while it runs and after it finishes. When the agent plans multi-step work, its plan shows above the message box and updates as steps finish.
Runs happen on the server. Closing the tab or losing your connection doesn't stop a run; open the chat again and it picks up where the stream left off.
Stop ends the run. The reply keeps the steps the agent finished before you stopped it.
Review tool activity
Open the Activity tab in the chat's side panel for the full record of the latest run:
- Where the run stands (Working, Done, or Stopped) and how long it has taken.
- The plan, with each step's state.
- Every tool call, with its duration. Open a call to see what was sent and what came back, or its error.
- What the run made, such as saved documents and files.
- The context it read in, and the input and output tokens it used.
Owners and admins can also see agents' tool requests across the workspace in Settings → Activity log: which tool was requested, by which agent, and when. The log doesn't include what was sent or returned.
When a tool call fails
A failed call shows as Couldn't followed by what the agent tried, and Failed in Activity, with the error. The agent sees the failure too and can try another way.
If a call failed because a connection's credentials were rejected, the step offers Reconnect for that connection. You can reconnect your own personal connections; for a shared connection, the step asks you to have a workspace admin reconnect it.
Uncertain outcomes
The chat shows each tool call as running, finished, or failed. A failed call didn't necessarily do nothing. When a request reached the other system but no confirmation came back, for example because it timed out or the run stopped partway through the call, the action may already have happened there. Its outcome is uncertain.
Adapt handles that the same way everywhere:
- It never retries a write whose outcome is uncertain, so an action can't happen twice without someone deciding it should.
- Agents are instructed to claim success only when a tool confirms it, and to say so when an outcome is uncertain.
- A reply to a surface, such as a Slack message, that may or may not have been posted is recorded as uncertain and isn't sent again.
- A sandbox whose creation is uncertain isn't created a second time. Check the sandbox provider, or wait for the sandbox's one-hour lifetime to end, before starting another conversation that needs it.
Before asking the agent to try a failed action again, check the other system for whether it happened.
How runs end
A run ends when the agent answers, when you stop it, or when it reaches the agent's time or step limit (see Set run limits). A run that stops early keeps the steps it finished, so your next message continues from them. The step in progress when it stopped is dropped, and the agent is told which tool calls were cut off, since those may still have taken effect.
Adapt checks permissions again before every tool call. If your role or the agent's tools change during a run, the next tool call is refused and you send a new message to continue. An action already sent to another system can't be taken back.
If the server restarts, a run that was between steps continues on another server process from its saved steps, without repeating a tool call. A run cut off in the middle of a step is marked interrupted and isn't replayed.
Long chats
When a chat grows past the agent's context budget, Adapt summarizes the older turns and keeps the most recent ones in full for the model. The full transcript stays in Adapt for you to read and for history search.
Who can see a chat
A web chat belongs to the person who started it. Nobody else can open it, including owners and admins. When an agent searches past chats, it searches only the chats of the person it's working for.
Some conversations aren't web chats:
- Surface threads, such as a Slack thread or a GitHub issue, are visible where they happened. The member who started one also sees it in their chat list. Other members who can see the agent can open it read-only in Adapt, subject to the surface's own access: a Slack thread opens only for people who can see it in Slack.
- Scheduled automation runs each get a conversation owned by the automation's creator. The agent's creator and workspace owners and admins can open them from the automation's run history. See Automations.
Surface and automation conversations can't be continued in web chat.
Rename, export, or delete a chat
Open the menu on the chat's name at the top of the chat:
- Rename changes the name in your chat list.
- Export conversation saves the conversation as Markdown, copied or downloaded. Include tool activity adds the tool calls. Attachment links in the export still require access to the conversation.
- Delete removes the chat and its history from your chat list.