Multi-Agent Systems: When to Split and When Not To

Q: True or false: 'In a multi-agent system, the more agents, the smarter and less error-prone it is.'

False。Errors compound along the agent chain — one small problem can derail the whole chain — and cost is ~15x. Whether to split depends on whether the task is breadth-parallel and valuable enough.

First, separate: workflow vs. agent, and whether to split into many

First, bust a conflation: Anthropic separates 'workflows' (orchestrating LLMs and tools along predefined code paths) from 'agents' (the LLM autonomously decides the flow and tools) — if a workflow solves it, don't reach for an autonomous agent; keep it simple first. Should you split into multiple agents? Orchestrator-workers is the workhorse pattern: a central LLM dynamically splits tasks to sub-agents and aggregates. Common topologies: network / supervisor / hierarchical.

Multi-agent fits 'breadth-first, parallelizable, beyond a single context window' high-value tasks (e.g., one agent researching from several angles at once); it's a poor fit for tightly-coupled, strongly-dependent tasks (split them and they fight each other). Systems like MetaGPT encode human SOPs and role division into a pipeline plus structured document hand-off, reducing cascading hallucination.

Multi-agent's two bills: 15x cost, and errors that compound

Splitting into multiple agents isn't free. Cost: multi-agent runs about 15x the tokens of a plain chat (single-agent ~4x) — don't split mindlessly for 'smarter.' The gains are two-sided too: Anthropic's internal eval had multi-agent beat single-agent by 90.2% (only on its internal research-style eval, not general benchmarks), but academia (e.g. Cemri et al. 2025) finds the gains on most benchmarks are actually small; and failure modes can be catalogued: MAST groups 14 failure modes into 'system design / inter-agent misalignment / task verification.'

The myth to remember most: 'more agents = smarter' is wrong — errors compound along the agent chain, and one small problem can derail the whole chain. That said, multi-agent has a signature trick: multi-agent debate (multiple LLM instances critiquing each other over rounds to converge) significantly improves reasoning accuracy and factuality. Three framework paradigms to compare: AutoGen (conversational), CrewAI / MetaGPT (role-task), LangGraph (graph).

⚠️Default to single-agent + workflow first; only split into multi-agent when the task is genuinely breadth-parallel and valuable enough to be worth 15x the tokens.

自测 · 学完检查一下

想真正动手做题、记进度、攒连胜？到互动课里练。

Per Anthropic's distinction, the key difference between a 'workflow' and an 'agent' is?

答案：A workflow orchestrates along predefined code paths; an agent has the LLM autonomously decide the flow

Workflow = predefined paths, agent = LLM autonomy; if a workflow solves it, don't reach for an autonomous agent — keep it simple first.

Which kind of task is best suited to splitting into multiple agents?

答案：Breadth-first, parallelizable, high-value tasks beyond a single context window (e.g., researching from several angles at once)

Multi-agent suits breadth-parallel work; tightly-coupled tasks fight each other when split, and multi-agent costs ~15x the tokens — it has to be worth it.

True or false: 'In a multi-agent system, the more agents, the smarter and less error-prone it is.'

答案：False

Errors compound along the agent chain — one small problem can derail the whole chain — and cost is ~15x. Whether to split depends on whether the task is breadth-parallel and valuable enough.