Your chat with an agent

One chat with each agent#

Each agent has one ongoing chat. Open the agent from the sidebar and write: there is no list of chats to manage. On other pages, the agent at the top right opens its chat beside the page. Messages, the agent's replies, and the results of its automations stay in that one chat, so you can pick up where you left off.

When other people have access to the agent, they use the same chat: everyone sees it and writes in it, and each message shows who wrote it. Each person has their own unread state, so a message you have read can still be new to a teammate. See Give people access.

The chat never resets. When it gets long, the agent compacts it: it keeps a summary of the earlier messages (people, decisions, preferences, work done, open questions) and the latest ones word for word, and a Chat compacted card marks where it happened; choose Show summary to read what it kept. Nothing is deleted. You can scroll back and search everything, and so can the agent: ask about something from weeks ago, a decision or a file, and it looks it up in the chat.

What you see while the agent works#

The chat reads like messaging a teammate: your messages and the agent's appear as chat bubbles. While the agent works, a row at the bottom of its reply shows its moving avatar and a short line saying what it is doing, such as the step it is on or Working…, with the time taken once the work runs long. While an approval waits for you, the line reads Waiting for your approval.

The tools the agent uses, its reasoning, and the other steps behind each reply are hidden. What needs you, or what the agent made, still shows: an approval request waiting above the composer, the agent's questions, cards to connect an account, the files it made, a notice when you run out of credits or a connected ChatGPT plan reaches its limit, the live browser while the agent uses it, and the Chat compacted card. A reply that ended with nothing to show reads Finished without a reply, and one that was stopped reads Stopped.

To see the steps, turn on Show the agent's steps in Settings → General, or ask the agent to turn it on for you. The chat then also shows the tools behind each reply, the agent's reasoning, and its plan and sub-agents, with its messages still in bubbles. The setting is yours alone: it does not change what a teammate sees in a shared chat. A reply describing an action is not proof that the action has finished; check the resulting record, or turn on the steps to read the tool result.

Send while it works#

You don't have to wait for a reply. Write while the agent works, and your message shows at the bottom of the chat, marked Read after the current step. The agent doesn't stop: it reads your message once the step it is on ends and carries on with it in mind, and what it says next appears below your message. Several messages sent during one step reach it together. While the agent works, the button beside the input is the stop button; on a computer, press Enter to send.

In a shared chat, a message a teammate sends while the agent works on someone else's request waits until that work ends, and then gets its own reply.

Write a useful brief#

Include four things when they matter:

Who will use this portal?

Our customers

Our partners

Our team

Submit

Next to its avatar, the agent asks who will use the portal. Customers is selected; partners and the team are other options.

1. Outcome: what you want at the end.
2. Source: which app, records, file, or connection to use.
3. Constraints: deadlines, audience, format, and changes to avoid.
4. Review point: what you want to approve before the agent continues.

Use the attached supplier offers to create a comparison by unit price, minimum order, and lead time. Keep currencies separate. Show the comparison before creating any purchase orders.

Answer clarification questions directly. If the agent chooses the wrong record, identify the correct one using its name and another distinguishing detail.

The agent learns as you go#

Tell the agent about its job while you work: its role, who it writes for, how to write, rules to follow. It saves what is meant to last to its own description, personality and knowledge, which you can see and edit in its settings, and says so in a short line. When you ask for something that repeats, such as a summary every Wednesday, it sets up an automation for you to approve. When you describe a long, step-by-step way of doing one job, it proposes a skill for you to approve instead of adding the steps to its knowledge. One-off requests stay one-off. On a specialist agent other people use, everyone with access can change its personality, knowledge, rules, skills and automations, unless its owner turned that off; only its owner or an admin changes the rest.

Find earlier work#

Use the search button in the chat's header, or press ⌘F (Ctrl+F on Windows and Linux). Before you type, it lists the messages people sent, newest first, so you can jump back to any of them. Type to find words in messages, the agent's replies, and file names across the whole chat; more results load as you scroll. Choose a result to scroll the chat to it.

On a wide screen, a line of short marks beside the chat stands for the messages people sent around where you are, however far back. Point at a mark to read the message and click it to go there.

Open the gear in the header and choose Files to see every file in the chat: what you uploaded and what the agent made. Keep durable facts in Context or a shared knowledge app as appropriate; do not rely on a long chat as the only record of a business decision.

Files and interruptions#

Attach relevant files to the request. The agent keeps its working files in its workspace. Use the stop button when the task needs to change. Stopping does not undo actions that already completed, and messages the agent has not read yet are answered next. For a small change, you often don't need to stop: send it, and the agent picks it up after its current step.

For unexpected failures, check approvals, access, and connections before retrying a write. See troubleshooting for a checklist that avoids duplicate changes.

Talk instead of typing#

When the input is empty, choose the voice waveform button to talk with your agent about CRM records, deals, tasks, or other connected apps. Your speech and the agent’s reply appear in the conversation like any other message. Voice stays connected as you navigate. Before your first call, review the credit rate; live audio uses workspace credits even with a connected ChatGPT subscription. See Voice conversations for controls, voice previews, billing, and microphone access.