Hierarchical Multi-Agent

When a task grows past a certain size, one agent holding all of it starts to wobble — plan, details and half-finished subtasks all compete for the same context window. A hierarchical multi-agent system borrows the fix organisations use: one agent owns the goal and hands narrow pieces of it to agents that only need to know their piece.

How the layers split the work
The arrangement is a tree, not a conversation among peers:

• A supervisor holds the goal, decomposes it, and decides which subordinate gets which subtask.
• Each worker gets a brief and a narrow toolset — never the whole plan.
• Results come back up as summaries; the supervisor accepts, retries, or re-plans.
• A worker can itself be a supervisor, which is what makes the pattern hierarchical rather than merely delegated.

What it buys, and costs
The real gain is context isolation: a worker's thousand lines of tool output are compressed to a paragraph before they reach the supervisor, so top-level reasoning stays clean over long runs. The cost is that every hand-off is lossy. A worker reporting success it never achieved sends the supervisor confidently down the wrong branch, two siblings can redo each other's work, and each layer multiplies token spend. LangGraph's supervisor graphs and AutoGen's nested chats implement this shape.

Why it matters
Flat multi-agent setups degrade as they grow, because everyone talks to everyone. A hierarchy gives autonomy a reporting line — the structure that keeps a large task legible to the agent answerable for it.