Skip to main content

Tool calling

Tool calling is how a Builder turns a user request into concrete actions. It selects Tasks (in the LLM sense, tools) and executes them to produce results.

How it works​

  1. Interpret the intent from the user’s message.
  2. Select a task (Action, Tool, Variant, Plan Generator, or Artifact).
  3. Provide inputs using the task’s schema.
  4. Execute and return the output in chat.

Composable by design​

Because tasks share a common input/output contract, a Builder can chain multiple tasks or delegate orchestration to a Plan Generator. This enables workflows like:

  • Finding a document.
  • Transforming it (e.g., Word → PDF).
  • Sending it via email.

Plan Generator vs. multiple tool calls​

Using a Plan Generator can be more efficient than issuing many individual tool calls in a loop. With multiple tool calls, the LLM must ingest each tool response before deciding the next call. That increases token usage, latency, and the amount of sensitive data that flows through the LLM.

By contrast, a Plan Generator combines the same tasks into an execution plan that runs like code (without passing each step back through the LLM). In practice, only the final output is sent back to the LLM, instead of every intermediate tool response. This improves:

  • Data privacy: intermediate outputs do not flow through the LLM.
  • Token consumption: fewer intermediate tool responses to ingest.
  • Execution time: the workflow runs as code without LLM analysis between steps.