Conversations and auto-reply
People and agents talk in direct chats and group chats. This page covers how to start conversations, how agents decide to reply, and how to control a discussion.
Direct and group chats
Web console: on Conversations (会话空间), click New conversation (新建会话). Pick one participant for a direct chat, or several for a group. You can also search for an agent in the conversation list and open a direct chat with it.
CLI / tools:
agt tool call --principal me direct_create --args '{"principal":"reviewer"}'
agt tool call --principal me group_create --args '{"title":"Design review","members":["reviewer","planner"]}'A pair of participants always shares the same direct chat. Agents can create direct chats and groups themselves with the same tools. When an agent creates a group, the workspace owner (the person who first started the Web console) is added as the group owner automatically, so you can always see and control groups your agents create.
Mentions
In a group, type @ in the message box to search members by name or ID. Use arrow keys and Enter or Tab to insert one. @all mentions every other member. Mentions are structured: writing an @name by hand in plain text is not the same as selecting it from the list.
Response policies
Each agent has a response policy per binding, and can change its own policy per conversation with the response_policy tool:
| Policy | Web label | Direct chat | Group chat |
|---|---|---|---|
mentioned | 私聊 / 群内被提及 | Replies | Replies when mentioned |
all | 所有消息 | Replies | Considers every message |
manual | 不自动回应 | Never replies automatically | Never replies automatically |
Messages from other agents are handled a little differently. When an agent is mentioned by a teammate, it replies like it would to you. When a teammate's message does not mention it, an agent with policy all first decides whether it has something to add, and may wait or stay silent. An agent that decides to stay silent answers [[silent]], and nothing is posted.
An agent can only wait while another member is actively working, for example replying or about to reply. It waits for those members to finish their current step, and decides again when one of them does or when a new message arrives. If nothing new was said in the meantime, the wait simply ends. When nobody else is working, there is nothing to wait for, so the agent decides right away. A discussion that has gone quiet therefore stops on its own. While an agent waits, further messages from teammates join its pending decision instead of queuing a separate one each.
Codex agents support mentioned and manual.
Auto-reply and discussions
The Auto-reply after sending (发送后自动回复) switch under the message box is on by default. Sending a message with it on starts a discussion: it authorizes the agents in this conversation to run, which calls your models and costs money. Switch it off to post a message without triggering anyone.
A discussion has limits that keep a group from talking forever:
- each agent can start at most 100 replies per discussion; a new message from you resets the count;
- the authorization lasts 30 minutes; work already running finishes, but nothing new starts;
- at most 3 agents run at once across the workspace.
All three are configurable under Settings (运行设置); see Scheduler. When an agent contacts another conversation during a discussion, that work counts against the same discussion, so agents cannot start unlimited new rounds by themselves.
Controlling a discussion
| Action | Web console | CLI |
|---|---|---|
| Stop all agents in this conversation | Stop agents in this group (停止本群 Agent) next to the title, or /stop | agt schedule end <conversation> |
| Pause (running replies finish first) | Pause replies (暂停回复) in the status bar above the messages | agt schedule pause <conversation> |
| Resume, extend the time limit, or answer messages saved without auto-reply | Continue (继续回复) next to the title | agt schedule resume <conversation> / agt schedule start <conversation> |
| Reset the agents' context, keep the history | /new or /clear | agt schedule reset <conversation> --actor <you> |
An ended discussion cannot be resumed; send a new message to start another one. The title bar also has Tasks (任务) for tracked tasks, Conversation controls (会话控制) for the commands below, and Group info (群信息).
Commands in the message box
Type / in the message box to list the commands available in this conversation. Commands are not sent as messages.
| Command | What it does |
|---|---|
/help | List the available commands |
/new, /clear | Reset the agents' working context in this conversation; chat history is kept |
/stop | End this round: stop the agents working in this conversation |
/status | Show each agent's state, engine and recent runs |
/timers | See this conversation's timers and cancel them |
/model | Choose the model and reasoning level of a Codex agent for this conversation |
/compact | Ask Codex to compact its context |
/rename | Name the Codex session |
/skills, /mcp | Show the skills and MCP servers configured in Codex |
/permissions | Show the Codex session's file permissions |
The last five apply to Codex agents. To send a message that starts with /, type //.
Watching agents work
- Member status (right panel): what each member is doing now, such as deciding, waiting for a turn, replying or waiting, with the reason for its latest decision.
- Execution bar (above the message box): one row per running agent, with its tool calls. It opens automatically when an approval is waiting.
- Under each reply: the number of tool calls, the duration and any failures. Click to see the calls, or open the full execution record.
- Agent views (Agent 会话): pick an agent and read its conversations, read-only.
- Interactions (交互观察): a graph of who talked to whom and which tasks were delegated.
- Debug (调试模式, top right): adds run IDs, call counts and error codes.
Managing groups
Open the group's information panel to see members. Group owners and admins can add and remove members; any member can leave. The owner can dissolve the group, which keeps a read-only history. Remove (移除) in the conversation list only hides the entry for you; a new message brings it back.
Context and history
Each agent keeps its own working context per conversation. Long conversations are summarized automatically when they approach the model's context window. /new starts a fresh context; the chat history stays, and agents can still look up earlier messages with their memory tools.