← Contents Volume VII · Lecture 16

Volume VII · Memory and Many Minds

16

One Agent Becomes Several

Every mental model so far in this volume, and most of Volume V, assumed one agent doing the reasoning and one agent doing the acting. The obvious next question — what if several agents work on the same problem — has a short answer and a long history. The short answer, stated in Anthropic's own disclosed numbers, is that it can help a great deal and it is never free. The long history is worth walking in order, because each stage corrected a real failure of the one before it.

The mental model

Multiple agents are not automatically better than one — they are a way to trade tokens for coverage or accuracy, and the trade only pays off when the task actually parallelizes or actually benefits from disagreement. Anthropic's own production system spent roughly 15× the tokens of a single chat interaction to win 90.2% against one strong agent working alone — a real, stated cost, not a vague gesture at one.

2023: brittle loops, no paper

AutoGPT and BabyAGI, both from March 2023, are worth naming first because they are what "multi-agent" meant in the popular imagination before any of the more disciplined work below existed — and worth flagging plainly as press- and community-sourced: neither has a peer-reviewed paper behind it. BabyAGI ran a loop of task creation, task execution via GPT-4, result storage, and task reprioritization, using a vector database to store task-result embeddings so later steps could retrieve earlier context. AutoGPT decomposed a user's goal into sub-tasks and, at each step, self-prompted through a "thoughts / reasoning / plan / criticism" structure, evaluating its own progress and calling external tools and APIs along the way. Both were widely reported, at the time, for a specific failure mode: looping or getting stuck with no external correction, carrying far less scaffolding discipline than the frameworks that followed.

AutoGen: formalizing agent-to-agent conversation

AutoGen (Wu et al., arXiv:2308.08155, Microsoft, 2023) is the first entry here with a proper framework behind it rather than a community project. Its unit is the "conversable agent" — a customizable wrapper around an LLM, a tool executor, a human, or some combination of the three, built to communicate with other conversable agents through automated chat messages. A multi-agent workflow, autonomous or with a human in the loop, is constructed by declaring which agents talk to which and under what conversation pattern — the structure is explicit and configured, not emergent the way AutoGPT's self-prompt loop was.

Multi-agent debate: disagreement as the mechanism

Multi-agent debate (Du et al., arXiv:2305.14325, ICML 2024) is narrower in scope than AutoGen but sharper in what it claims. Multiple instances — copies of the same LLM — independently propose an answer and its reasoning to the same problem, are then shown each other's responses, and revise their own answer over several rounds, converging toward a shared final answer. The procedure is a black-box wrapper: no fine-tuning, the same prompts and procedure applied across every task the paper evaluates. The reported effect is improved accuracy on mathematical and strategic-reasoning tasks and reduced hallucination and factual error compared to a single instance generating alone — the benefit here comes specifically from disagreement and revision, not from splitting a task into parallel pieces.

Two different reasons to use more than one agent

AutoGen's conversable agents are a general-purpose way to structure who talks to whom. Multi-agent debate is a specific technique for one purpose — using disagreement among copies of the same model to catch errors a single pass would miss. Neither is "the" multi-agent pattern; they answer different questions about what multiple agents are for.

Anthropic's orchestrator-worker system: the real cost, disclosed

Anthropic's own multi-agent research system (anthropic.com/engineering/multi-agent- research-system) is the primary, engineering-blog-sourced case here, and it is the one that puts a number on the tradeoff the rest of this lecture has only gestured at. A "lead agent" analyzes the incoming query and forms a research plan, then spawns parallel subagents, each handed an explicit objective, an expected output format, tool guidance, and clear task boundaries. Subagents run independently — using tools such as web search — and report their findings back to the lead agent, which synthesizes across all of them and decides whether another research cycle is needed before answering.

ORCHESTRATOR-WORKER, WITH ITS TOKEN COST LABELED lead agent ~4× tokens of a single chat turn subagent 1 own objective subagent 2 own objective subagent 3 own objective findings reported back for synthesis · whole system ~15× tokens of a single chat turn
Figure 16a The lead agent plans and spawns; subagents run independently and report back; the lead agent synthesizes and decides whether another round is needed. The two token multipliers are Anthropic's own disclosed figures, not estimates.

The disclosed numbers are specific enough to quote directly rather than paraphrase. Agents in general typically use roughly 4× more tokens than a plain chat interaction; multi-agent systems use roughly 15× more tokens than chat. On Anthropic's internal research evaluation, a multi-agent system — an Opus 4 lead agent coordinating Sonnet 4 subagents — outperformed a single Opus 4 agent working alone by 90.2%. Token usage alone explained 80% of the variance in performance across the configurations Anthropic tested — meaning most of what separated a strong configuration from a weak one was simply how much token budget it spent, not some other architectural choice. Parallel tool calling, within and across subagents, cut research time by up to 90% on some multi-part queries.

Anthropic's disclosed multi-agent numbers, stated plainly
QuantityDisclosed figure
Tokens used by an agent vs. a plain chat turn~4×
Tokens used by the whole multi-agent system vs. a plain chat turn~15×
Accuracy win: multi-agent (Opus 4 lead + Sonnet 4 subagents) vs. single Opus 4 agent90.2% better, on Anthropic's internal research eval
Variance in performance explained by token usage alone80%
Research-time reduction from parallel tool calling, some multi-part queriesup to 90%
Plate 16.1 Anthropic's own token multipliers from Figure 16a, turned into an actual token count and dollar figure for a chat turn you size yourself — the 90.2% accuracy win came bundled with exactly this bill.

The line from 2023 to production

Read across the four systems in order and a throughline appears. AutoGPT and BabyAGI showed multiple steps and self-directed looping were possible at all, with essentially no governance over when to stop or how to divide labor. AutoGen supplied the missing structure — an explicit, configurable topology of who talks to whom. Multi-agent debate showed a narrower, specific payoff from disagreement among copies of one model. Anthropic's orchestrator-worker system is the production synthesis of the useful pieces from all three — explicit task boundaries per subagent, parallel execution, a synthesis step — plus something none of the earlier work disclosed: an honest number for what the coordination costs.

Read the primary source