Agent Tool Configuration

Available Tools

Select tools as needed in the Tools section of the Agent Editor page. Selected tools apply to all sessions of that agent.

Tool	Purpose	Typical Use Cases
`bash`	Execute shell commands inside the sandbox	Run scripts, install packages, invoke CLI tools
`write`	Write content to files	Generate reports, save intermediate results
`read`	Read file contents	Inspect mounted CSVs, logs, or configuration files
`edit`	Modify files via string replacement	Safer for small changes than full-file overwrites
`glob`	Find files using glob patterns	Locate target files, e.g., `*.csv`
`grep`	Search text within files	Debug logs, locate keywords
`web_search`	Search the internet and return the title, link, and content summary of relevant pages	Look up real-time information (such as weather and news) and external resources
`web_fetch`	Read the main content of a given web page and convert it to text	Read pages in depth after a search, extract page content
`mark_artifacts`	Mark files produced by the agent as result artifacts and return their `file_id`	Download or display the final deliverables on the business side

web_search and web_fetch are billed separately: web_search is billed per call (CNY 0.03 per call); web_fetch is free of charge for a limited time—future pricing is subject to the console. For billing details, see Billing.

Execution Mechanism

All built-in tools execute within the session-bound runtime environment. Files created and dependencies installed by these tools persist for the duration of the session. The agent autonomously decides which tool to invoke based on system prompts and the current message. Tool invocations are visible in real time via the Events panel.

Approval Policy

In the Agent Editor's Tools section, expand a tool to set its own approval policy, controlling whether a call requires human confirmation before it runs:

Policy	Behavior
Always allow	Default. The agent runs the tool directly without interrupting the flow
Always ask	Pauses before each call and waits for the user to confirm in the session before continuing

For high-risk tools (such as bash), set Always ask to keep a human gate on sensitive operations while retaining automation elsewhere. When a tool is set to Always ask, the agent pauses and waits for confirmation each time it calls that tool during a session. When integrating via the API, see Session Event Stream for how to receive and respond to approval requests in the event stream.

Scope of Approval

Tool approval is supported only for the main agent and is not supported for subagents.

Client-side tools (function / custom, etc.) do not participate in this approval mechanism; their execution is controlled by the client.

In addition, a session locks in the Agent version at creation time. Changing an Agent's approval policy does not retroactively affect existing sessions — you must create a new session for the new configuration to take effect.

Configuring Approval Policy via the API

Beyond the console, you can declare the approval policy through the tools field when creating or updating an Agent. Set the toolkit-level default in the default_config of a builtin_toolkit entry, and override it per tool in configs[]:

{
  "tools": [
    {
      "type": "builtin_toolkit",
      "default_config": {
        "enabled": true,
        "permission_policy": {"type": "always_allow"}
      },
      "configs": [
        {
          "name": "bash",
          "enabled": true,
          "permission_policy": {"type": "always_ask"}
        }
      ]
    }
  ]
}

Configuration rules:

permission_policy only accepts the object form {"type": "always_allow"} or {"type": "always_ask"}. Passing a string returns a parameter error.
A per-tool policy in configs[] overrides the same field in default_config; a tool not listed in configs[] inherits default_config.
The approval policy for MCP tools is configured in the default_config / configs of the tools[] entry whose type is mcp_toolkit — not in mcp_servers[].
When omitted entirely, the default policy is always_allow.

For full field definitions, see the Create Agent API.

Next Steps

Define an Agent: Select required tools in the Agent Editor.
Agent MCP: Extend functionality by integrating additional external tools via the MCP protocol.
Start a Session: Observe tool invocations in action during a session.

​Available Tools

​Execution Mechanism

​Approval Policy

​Scope of Approval

​Configuring Approval Policy via the API

​Next Steps

Available Tools

Execution Mechanism

Approval Policy

Scope of Approval

Configuring Approval Policy via the API

Next Steps