Tools: Agent

The agent tool allows you to create sophisticated multi-LLM workflows by enabling one interlocutor to call another as a tool. The “agent” interlocutor receives the query from the “caller” with no other context, processes it, and returns its response as the tool’s output.

This is a powerful way to separate concerns. You can create specialized agents and then compose them into more complex systems. For an excellent overview of the philosophy behind this approach, see Anthropic’s blog post on their multi-agent research system.

Configuration

To use the agent tool, you must have at least two interlocutors defined. In the configuration for one interlocutor, you add an agent tool that points to the name of the other.

• agent: (Required) The name of the interlocutor to be called as an agent.
• name: An optional name for the tool.
• usage: A string, file:, or exec: URI providing instructions for the calling LLM on when and how to use this agent.
• raw_output: A boolean value. Normally an agent’s output will be sanitized, so that raw tool call results are not visible to the interlocutor who called the agent. Setting raw_output to true puts the full output from the agent into the main interlocutor’s tool call results.

Example Configuration

In this setup, Kirk is the main interlocutor. He has a tool named communicator which, when used, will call the Spock interlocutor. Spock has his own set of tools.

interlocutors:
  - name: Kirk
    prompt: You are Captain Kirk. You are bold and decisive.
    tools:
      - agent: Spock
        name: communicator
        usage: Use this to contact Spock for logical analysis and advice.

  - name: Spock
    prompt: >
      You are Mr. Spock. You respond with pure logic, suppressing all
      emotion.
    tools:
      - exec: bc

Example Conversation

Using the configuration above, Captain Kirk can delegate complex analysis to Spock.

We've encountered an alien vessel of unknown origin. It is not responding to
hails. What is the logical course of action?

:::Kirk

This situation requires careful analysis. I will consult my science officer.

<tool-call with="communicator">
<arguments>
<content>
┆Alien vessel, unknown origin, unresponsive. Propose logical course of action.
</content>
</arguments>
<results>
<result type="text">
┆Insufficient data. Recommend passive scans to gather
┆information on their technological capabilities before initiating
┆further contact. Avoid any action that could be perceived as
┆hostile.
</result>
</results>
</tool-call>

A logical approach. We will proceed with passive scans.

:::

Recursion

Agents can call other agents, and those agents can call back. There’s no built-in limit on recursion depth, so be thoughtful about your configurations. An agent that calls itself (or creates a cycle) will keep going until it decides to stop or hits a token limit.

For example, Spock could have a tool that calls Kirk:

interlocutors:
  - name: Kirk
    prompt: You are Captain Kirk.
    tools:
      - agent: Spock
        name: consult_spock

  - name: Spock
    prompt: You are Mr. Spock.
    tools:
      - agent: Kirk
        name: consult_captain

This is powerful for workflows where agents need to collaborate, but it requires clear instructions in the prompts about when to delegate and when to answer directly.