Tool

Tool

Tools let agents interact with the outside world instead of only producing text.

Two tool categories

Functional tools

These are the tools controlled by allowed_tools:

• builtin tools such as bash, web_search, and web_reader
• custom BaseTool implementations
• nested agents exposed through Agent.as_tool()

System tools

These are runtime-owned tools and are not filtered by allowed_tools:

• SkillTool
• scheduler runtime tools such as spawn_child_agent, fork_child_agent, and sleep_and_wait

Stable contract

Public tools implement BaseTool with:

• name
• description
• get_parameters()
• execute(...)

Plain tools should depend on ToolContext, not on agent-internal runtime types.

Result semantics

Use:

• ToolResult.success(...)
• ToolResult.failed(...)
• ToolResult.aborted(...)
• ToolResult.denied(...)

This keeps tool outcomes explicit for the runtime and for the model’s next step.

Builtin tools

The builtin surface currently includes:

• bash
• bash_process
• web_search
• web_reader
• memory_retrieval

You can treat them as defaults and layer your own tools alongside them.