Chatting with an agent

An agent’s page opens on its Chat tab. The threads are on the left, the open conversation is in the middle, and the configuration panel is on the right. Open an agent from the Agents page.

Every member of the organization can chat with every agent. Chat is for signed-in members; an organization key can’t open a thread. To reach an agent from your own code, see Talking to agents.

Start a thread

1. Open the agent. A new thread is ready when no thread is selected.
2. Type in the Message box. Enter sends, and Shift+Enter starts a new line.
3. The first line of your first message becomes the thread’s title, up to 80 characters.

Press New thread at the top of the chat to start another. While the agent replies, the send button turns into Stop. The badges in the message box show the agent’s provider and model.

What a reply shows

• Text, rendered as Markdown.
• Thinking, folded, when the model thinks before it replies. Press it to read.
• Tool calls, one row each with the tool’s name and a status: Preparing, Running, Waiting for approval, Done, Failed or Denied. Press a row to see its input and output, or the error.

Approve or deny

When the agent calls an action that asks first, the call shows as a card marked Waiting for approval, with the action’s name and its input. Press Approve to run it, or Deny to refuse. The agent continues its reply either way.

Whoever has the thread open decides. A thread a workflow run started also shows its waiting calls here, and a decision in chat counts for the run. See Grants and approvals.

Manage threads

• Search threads filters the list by title.
• The menu beside a thread has Rename and Delete. A deleted thread’s messages are gone for good.
• A thread you write in moves to the top of the list.
• The button at the top left of the chat hides or shows the list.

Threads from workflow runs

Each time an Agent step runs, it opens a new thread on its agent. The title names the workflow, the first eight characters of the run’s id and the step’s title, separated by dots. The workflow icon beside it opens the run. The thread holds the task the step sent and the agent’s whole reply.

When an agent has run threads, Hide run threads appears above the list with their count. Your choice is remembered in this browser for every agent.

You can write in a run thread after the run. The agent answers as in any chat, and the run doesn’t change.

Memory and files

An agent keeps two kinds of memory.

• This thread: notes the agent keeps for one conversation. No other thread sees them, and a new thread starts empty.
• Shared across threads: facts every thread reads and writes, such as standing preferences or policies. When you ask the agent to remember something beyond this conversation, it saves it there.

Files the agent writes go to its workspace, which all of its threads share. The agent can give you a link to one of its files. Anyone with the link can open it for 7 days. Deleting the agent deletes its files.

The configuration panel

The panel beside the chat shows the agent as saved:

• Apps: each grant with its account, how many actions it allows, and how many ask first.
• Skills and Databases: what it may use, with writable databases marked.
• Workspace tools, Browser tools and Memory: built in and always on.
• Instructions: rendered the way the agent reads them, with today’s date and its apps filled in.

Each Edit link opens the Editor tab. The button beside the tabs hides or shows the panel.

When the agent changes

A saved change to the agent reaches its existing threads too. The agent reads its new instructions, model and grants at your next message in each thread. See Creating an agent.

If someone deletes the agent while you have it open, the page says so and the chat closes.