Agent Recursion

When AI agents start using tools, the context window fills up fast. Each run_command or read_file tool call dumps a lot of text that quickly loses relevance in subsequent steps.

I’ve been experimenting with the idea of “recursive agents” to mitigate this issue. The idea is simple: An agent can kick off new instances of itself to handle specific subtasks.

The key part is that, when a child agent finishes, it doesn’t send its entire operational log or full tool output back to the parent; instead, it returns only a brief summary of the results.

This does two important things:

1. Keeps the parent agent’s context window usage down.
2. Helps the parent agent stay focused on the overall objective.

I’ve implemented this idea here: https://github.com/icholy/sloppy. The agent is simply provided with a run_agent tool.

{
  "name": "run_agent",
  "description": "Delegates a sub-task to a new child agent. Plan and break down work; use new, uniquely named agents for distinct sub-tasks or repetitive instances (min 3). Don't delegate the entire core task.",
  "parameters": {
    "type": "object",
    "properties": {
      "name": {
        "type": "string",
        "description": "A unique, short, and descriptive name for the child agent instance."
      },
      "prompt": {
        "type": "string",
        "description": "Instructions for the new child agent, or follow-up questions for a previously created agent."
      }
    },
    "required": ["name", "prompt"]
  }
}

Note: The description fields included here are abbreviated.