Vol. XXV · Tools and Agents Plate 7 of 9multi-agent systems, orchestrator, sub-agents, parallelism, tokens, coordination, research agentsCE

Many Agents.

Divide the search among helpers with their own windows. Pay for every one of them.

A single agent works through a task in one context, one step at a time. A multi-agent system has a lead agent — an orchestrator — that plans, spawns sub-agents with focused instructions, lets them work in parallel in separate contexts, and merges what they return. The separate contexts are the point as much as the parallelism: each sub-agent can read a large amount and return only a condensed result, so the lead’s window holds conclusions rather than raw material (plate XXV.4).

Anthropic described such a system for research questions in June 2025: it outperformed a single agent with the same model by 90% on their internal research evaluation, used about 15 times the tokens of a chat (single agents about 4 times), and found that token usage alone explained 80% of the variance in performance on a browsing benchmark. The workshop models a task of fixed total work split among m sub-agents, with a per-agent coordination overhead, and computes wall-clock time and tokens.

Below: wall-clock time and total tokens against the number of sub-agents, for the chosen task size and overhead; readouts at the chosen number.

What you will learn
Workshop
A task of W = 600 agent-minutes of reading, divisible evenly; planning 5 minutes; each sub-agent adds c minutes of the lead’s coordination (instructions, reading its report) and a fixed s = 20 000 tokens of setup and o = 5 000 of report; tokens read at r = 5 000 per minute. A model of the trade-off, not a measurement; Anthropic’s figures are quoted from their report.

Breadth for tokens

With 2 minutes of coordination per sub-agent, eight sub-agents finish the 600-minute task in 96 minutes instead of over ten hours, and the fastest split is about 17 agents, beyond which coordination costs more than parallelism saves. Tokens only rise: setup and reports for eight agents cost 1.06 times a single agent’s tokens here — and in practice far more, because each sub-agent explores on its own, which is how Anthropic’s system came to use about 15 times the tokens of a chat.

The gains are real for tasks that divide into independent parts — surveying many sources, checking many files, trying several approaches — and small or negative for tasks whose steps depend on each other, where sub-agents duplicate work or act on stale assumptions. Anthropic reported that much of their engineering went into the orchestrator’s instructions: how to split a task, how much effort each part deserves, and how to describe the boundaries so sub-agents do not overlap. Multi-agent systems trade money for time and breadth; whether that trade is worth making depends on the task.

Sub-agents also isolate failure: one that wanders down a dead end costs its own tokens but returns only a short report, leaving the lead’s context clean.

Terms

Orchestrator
The lead agent that plans, delegates and combines.
Sub-agent
An agent given a focused part of a task in its own context.
Coordination overhead
The lead’s cost of instructing and reading each sub-agent.
Parallelism
Working on independent parts at the same time.
Multi-agent systems buy speed and breadth with tokens; they pay off when a task divides into independent parts.

Sources

  1. Anthropic Engineering, “How we built our multi-agent research system”, 13 June 2025.
  2. Wu et al., “AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation”, arXiv:2308.08155, 2023.