Skip to content

18.04 Pillar 2: Hierarchical Delegation & Sub-Agents

The second critical pillar of a Deep Agent is Hierarchical Delegation—the ability for the primary agent to spawn instances of itself to tackle specialized sub-tasks.

The "Handyman" Analogy

If you need to fix a leaky skylight on a high ceiling, and you lack the tools or the expertise, you don't climb a ladder and attempt it yourself. You hire a specialized craftsman (a handyman).

You provide the handyman with a high-level goal ("Fix the leak"). The handyman arrives with: 1. Their own expertise (A specialized System Prompt). 2. Their own ladder and tools (A restricted, specialized Toolset).

The handyman climbs the ladder and works in isolation. You don't need to monitor every single screw they turn. When they finish, they come down and give you a final report ("It's fixed").

The Technical Implementation: Sub-Agents

This exact concept applies to Deep Agents through Sub-Agents.

When a Deep Agent realizes a sub-task is complex, rather than executing it directly and polluting its own context window with intermediate errors and file searches, it spawns a Sub-Agent.

How it Works:

  1. The Main Agent writes a specific, focused prompt for the sub-task.
  2. The Sub-Agent is instantiated. It has its own unique system prompt (e.g., "You are a dedicated authentication researcher...") and its own isolated tool calling loop.
  3. The Sub-Agent executes its task locally, potentially failing, retrying, and reading dozens of files.
  4. When finished, the Sub-Agent dies and returns only the final, condensed answer back to the Main Agent.
graph TD
    A[Main Agent Context Window] -->|Spawns| B(Sub-Agent 1: Explore Auth Patterns)
    A -->|Spawns| C(Sub-Agent 2: Write Unit Tests)

    subgraph Isolated Context
        B <--> D((Web Search Tools))
        B <--> E((File Read Tools))
    end

    B -->|Returns Final Summary| A
    C -->|Returns Test Results| A

Why Sub-Agents are Crucial

  • Context Isolation: By keeping the noisy, intermediate tool calls inside the Sub-Agent, the Main Agent's context window remains clean, lean, and focused on the high-level objective.
  • Parallel Execution: Because Sub-Agents are isolated, the Main Agent can spawn multiple Sub-Agents concurrently, drastically reducing the completion time for massive tasks.