Quick start

This guide takes you from a fresh install to a lead agent that builds its own team. It uses the Web console; the CLI section at the end shows the same steps in a terminal.

Console language

The console is in English and Chinese and follows your browser's language; switch it with the language menu at the top right. This guide names each control in English, with the Chinese label in parentheses, for example Resources (资源库).

Before you start, install agt and start the console with agt web. Open the full link it prints.

1. Connect a model

Open Resources (资源库) in the left sidebar and click New resource (新建资源).

1. Create a Provider: a name, your endpoint's base URL (for example https://api.openai.com/v1) and the API key.
2. Create a Model on that provider. Click Fetch model list (获取模型列表) to choose from the models the provider offers, or type the name. Pick the API your provider supports for that model: Chat Completions or Responses. Keep tool calling enabled.

See Connect models for every option.

2. Choose how tool calls are approved

By default, every tool call that changes something or reaches outside agt waits for approval; reading agt's own data, such as message history or tasks, does not. Open Settings (运行设置) → Tool approvals (工具权限审批), and set Native approval policy (Native 权限审核策略):

Mode	When to use it
Model review, doubtful calls to you (模型审核，异常转人工)	A reviewer model approves routine calls and passes doubtful ones to you. Select the model from step 1 under Reviewer model (审核模型).
Manual (人工审批)	You approve each call yourself. Safest, but the lead makes many calls.
Allow all (自动放行)	No approvals. Only for experiments in a disposable home.

Without a reviewer model, model review falls back to manual approval. Pending requests appear above the message box in the conversation, where you can approve or deny them. See Permissions.

3. Create the lead

Open Agent team (Agent 团队) and click Create agent (创建 Agent):

• Name / identity ID (资源名称 / 身份 ID): lead
• Engine (执行方式): Native (AGT 内置), with the model from step 1
• Role prompt (角色 Prompt): Use an existing prompt (使用已有 Prompt), then default/lead
• Built-in tools (框架内置工具): tick Resource management (资源管理) and Task collaboration (任务协作)
• Response policy (响应策略): keep Direct chats and mentions (私聊 / 群内被提及)

Saving creates the agent definition, its chat identity and the binding that enables it, so the lead is ready to talk to.

4. Give it a goal

Click New conversation (新建会话), choose a direct chat with lead, and describe what you want, for example:

Set up a two-person review team for this repository: one reviewer focused on correctness and one on readability. Put them in a group with me and have them review the latest change.

Keep Auto-reply after sending (发送后自动回复) switched on. Sending authorizes the agents to run, which calls your model and costs money.

The lead checks existing resources, writes role prompts, creates and enables each teammate, creates the group and briefs the team. While it works:

• The bar above the message box shows each running agent and its tool calls.
• Pending approvals appear there too, if you chose manual approval.
• Agent team lists the new teammates as they are created. Click one to see its configuration, status and recent runs.
• Turn on Debug (调试模式) in the top-right corner for run IDs and error codes.

5. Let it build tools

With a Codex agent available, the lead can also build tools. Ask for something that needs a tool that does not exist yet, for example a small MCP server for a board game or a data source. The lead delegates the code to a Codex agent, registers the result as an MCP server, tests that it starts, and attaches it to the agents that need it. See Build a team with a lead.

Doing the same from the CLI

Everything the console does goes through the agt CLI. The example below creates a reviewer agent by hand and talks to it; the lead does these same steps through its tools.

export MODEL_API_KEY=sk-...
agt provider create main --base-url https://api.openai.com/v1 --api-key-env MODEL_API_KEY
agt model create fast --provider main --model gpt-5-mini --api-format responses --tool-call true
agt prompt create reviewer --text 'You review code changes. Stay silent with [[silent]] when you have nothing to add.'
agt agent create reviewer --model fast --prompt reviewer

# Chat identities: yourself and the agent, then enable the agent
agt im principal create me --name Me
agt tool call --principal me --manager principal_create --args '{"id":"reviewer","name":"Reviewer","kind":"agent"}'
agt tool call --principal me --manager agent_bind --args '{"principal":"reviewer","agent":"reviewer"}'

# Open a direct chat and send a message
agt tool call --principal me direct_create --args '{"principal":"reviewer"}'
agt tool call --principal me message_send --args '{"conversation":"<conversation id>","body":"Please review src/main.rs","request_key":"hello-1"}'

# Authorize the discussion and run the scheduler
agt schedule start <conversation id>
agt schedule run --watch

In a second terminal, read the reply with agt tool call --principal me message_history --args '{"conversation":"<conversation id>"}'. The module pages list every command and parameter.

Next steps

• Core concepts: resources, identities, discussions and tasks
• Create agents: prompts, skills, MCP servers and built-in tools
• Troubleshooting: what to check when an agent does not reply