Multi-Agent Systems: When You Actually Need More Than One
A multi-agent system splits a task across several AI agents that each hold their own context. It genuinely helps on breadth-first, parallel, read-heavy work like research, where Anthropic measured a large gain over a single agent, but at roughly fifteen times the token cost. It hurts on anything needing tight coordination or shared decisions, where parallel agents make conflicting assumptions. The honest default is a single strong agent with good tools; reach for multiple agents only when the task is truly parallelizable and its value clears the cost.
Open any AI feed right now and someone is announcing a swarm. A team of agents, a network of specialists, a mesh of bots negotiating with each other. It looks like the future. Most of it should be a single agent with better tools, and the people with the most data on this quietly agree.
Let me give you the real numbers, from both sides, so you can decide like an engineer instead of a hype-watcher.
The pitch, and the price tag
Anthropic built a multi-agent research system and published what it measured. The upside is real, and so is the bill:
Read that third number again, because it is the one that matters. When roughly four-fifths of the improvement is explained by how many tokens you spent, the "multi-agent magic" is largely just more compute applied to a problem that could absorb it. That is a genuinely useful thing, on the right task. It is not intelligence emerging from collaboration. It is a bigger search, run in parallel.
Where more agents genuinely win
The task that produced those numbers is the tell: find every board member across a large set of companies. That is breadth-first, read-heavy, and embarrassingly parallel, hundreds of independent lookups with no dependencies between them. Multiple agents shine exactly here:
- Research and search across many independent sources.
- Breadth-first questions with many directions to pursue at once.
- Work whose total information will not fit in one context window.
- Cleanly separable subtasks where one agent never needs another's half-finished thinking.
Anthropic is refreshingly blunt about the flip side: multi-agent is a poor fit for work that needs all agents to share context or that has many dependencies between steps, and it names most coding as exactly that kind of work.
Where more agents multiply your problems
Cognition, the team behind Devin, wrote the counterargument, titled bluntly "Don't Build Multi-Agents." Their point is sharp: when parallel agents cannot see each other's in-progress decisions, they proceed on conflicting assumptions, and the pieces do not fit. Their example is one subagent building a game background in a Mario style while another builds the character in a different style, leaving a final agent to stitch together two things that were never going to match. By 2026 they had refined the position rather than reversed it: multi-agent works best when the writes stay single-threaded and the extra agents add intelligence, like review or routing, rather than taking conflicting actions.
And the neutral evidence backs the caution. A Berkeley-led benchmark study of why multi-agent systems fail found that their gains over single agents are often minimal, and sorted the failures into three families: flawed specification of the system or task, misalignment between the agents, and weak verification of the final result. That is the honest counterweight to any single impressive demo.
So how do you actually decide
Start with one agent. The average multi-agent system does not beat a good single one, so complexity is a cost you have to justify, not a default you reach for.
Add agents only when all of this is true: the task is genuinely parallelizable and read-heavy, the subtasks are cleanly separable, and the value clears the roughly fifteen-times token premium. When you do, use the orchestrator-worker shape, one lead that plans, delegates to parallel subagents, and synthesizes, and keep the actual writes single-threaded so two agents can never take conflicting actions. Swarms of agents freely negotiating make a great screenshot and a terrible production system.
Three things worth remembering
- More agents multiply cost and failure surface at the same time. You pay many times the tokens for a system with more distinct ways to break. The upside only shows up on the narrow class of tasks where the parallelism is real.
- More agents is not more intelligence. When most of the measured gain is just token spend, adding agents to a task that cannot absorb parallel search adds noise, not IQ.
- The winning move is usually a better single agent with tools. Both the biggest proponent and the loudest skeptic converge here: single agent by default, orchestrator-worker for parallel research, never parallel writers.
How we build
We build AI agents and agentic workflow automation for clients, and we almost always start with one strong, well-instrumented agent and prove it with evaluation before anyone says the word "swarm." Multi-agent is a tool for a specific shape of problem, not a maturity level. If someone is selling you a mesh of agents for a job one agent could do, that is a cost, not an upgrade. Talk to us and we will tell you honestly which one your problem needs.
References
- Anthropic, How we built our multi-agent research system: the 90.2 percent gain, the 15x token cost, and the 80 percent variance figure.
- Cognition, Don't Build Multi-Agents: why parallel agents make conflicting decisions.
- Why Do Multi-Agent LLM Systems Fail? (Cemri et al.): the failure taxonomy and the "often minimal" gains finding.
Frequently asked questions
It is a setup where several AI agents, each with its own context window, divide a task and coordinate, rather than one agent handling everything in a single loop. The most reliable shape is orchestrator-worker: a lead agent breaks the task into pieces, delegates them to subagents that run in parallel, and synthesizes the results.
Sometimes, on the right task. Anthropic reported a multi-agent research system beating a single agent by about 90 percent on an internal research eval, but it used roughly fifteen times the tokens of a chat, and about 80 percent of the gain tracked token spend, meaning you are largely paying for more parallel search. A neutral benchmark study found the average multi-agent system's gains are often minimal, so it is not a free upgrade.
When the task is genuinely parallelizable and read-heavy, like researching many independent sources, when subtasks are cleanly separable so agents do not need each other's intermediate decisions, or when the work exceeds a single context window, and the value of the task justifies the much higher token cost. For most coding and anything needing shared state, a single agent is better.
Mostly coordination. When parallel agents cannot see each other's in-progress decisions, they proceed on conflicting assumptions and their work does not fit together. A benchmark taxonomy grouped the failures into three buckets: flawed system or task specification, misalignment between agents, and weak verification of the result. More agents means more surface area for exactly these failures.
Building something with AI agents?
We design and ship production AI agents for teams in fintech, crypto, and beyond.
Book a call