Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

2.9 KB
Raw Blame

Agent tools

Tools that act on the agent loop itself, its todo list, or the user's stored preferences, rather than on files, memory, or the network. crate::tools re-exports everything here (tools/mod.rs re-exports crate::agent::tools), and tools::ops registers them into the tool catalog; nothing outside this crate should construct them directly.

Contents

Tool File Wire name (if different)
AskClarificationTool tinyagents_harness_tools ask_user_clarification
plan_exit_tool plan_exit.rs plan_exit
RememberPreferenceTool remember_preference.rs remember_preference
SavePreferenceTool save_preference.rs save_preference
RunWorkflowTool, AwaitWorkflowTool run_workflow.rs run_workflow, await_workflow
TodoTool todo.rs todo

RunWorkflowTool and AwaitWorkflowTool are compiled in only with the skills feature; a build without it omits both tools from the catalog instead of registering them as disabled.

Key types

  • AskClarificationTool returns the question text as its own tool output; it does not pause anything by itself. The pause happens because callers list ask_user_clarification in the harness seam's early_exit_tools (tinyagents::TurnSpec::run), and a successful call there steers the loop to Pause instead of feeding the result back to the model. Any caller that exposes this tool without that listing gets a tool that answers its own question, since the model reads a normal success and keeps going.
  • plan_exit_tool configures TinyAgents' plan-exit tool with OpenHuman's thread mode and pending-review stores. TinyAgents owns the tool behavior and PLAN_EXIT_MARKER.
  • RememberPreferenceTool pins an explicit (class, key, value) preference into the user_profile memory namespace. SavePreferenceTool stores a free-form preference into either the general or situational lane. RunWorkflowTool and AwaitWorkflowTool spawn a crate::skills::runtime workflow run and wait on its outcome.
  • TodoTool is the model-facing side of the session's todo list; TinyAgents owns its schema and store operations, while the host adapter selects the thread key and publishes the product event.

Where next

  • crate::tools for the shared Tool trait, ToolSpec, and how these tools join the rest of the catalog.
  • crate::agent::orchestration for the sub-agent spawn tools, which sit next to these conceptually but live separately because they carry more lifecycle state.
  • tinyagents_harness_tools::todos for the todo store and tool behavior.

Further reading