Multi-Agent Systems
Several AI agents working together on one problem, each with a role — powerful in demos, awkward in production.
When not to use it
- When a workflow would do. Fixed steps with a model call at each one is the right answer far more often than a team of agents, and it can be debugged.
- When the sub-tasks depend on each other. Agents coordinate badly; sequential dependencies remove the only real benefit, which is parallelism.
- When cost or latency matter. Every agent is a full inference loop. Three agents is roughly three times the bill, plus the messages between them.
Reach for something else instead
- A single agent with tools — usually the same capability with one context and one place to look when it breaks.
- A workflow — explicit steps, deterministic control flow, model calls where judgement is genuinely needed.
- One model, multiple passes — draft then critique in sequence, which captures most of the critic benefit without a second agent's overhead.
Read more on the blog
- Multi-agent AI gets worse as you add agentsOrchestrator steering accuracy falls from around 60% with three agents to about 21% with ten. Coordination is not free, and most teams reaching for multi-agent should fix their single agent first.
- Why AI agents fail: the seven failure modesGartner predicts over 40% of agentic AI projects will be canceled by 2027. The failures follow patterns, seven of them. The taxonomy: what breaks, why, which real incident proved it, and which control would have prevented it.
- When not to use an agentGPT-3.5 inside a structured workflow scored 95.1% on a coding benchmark. GPT-4 running free scored 67%. The structure was worth more than two generations of model improvement, and most tasks called agentic do not need an agent.
- Agent permissions: the question nobody asks until afterwardsEighty percent of organisations running agents say those agents have taken unintended actions. One in five has had a security incident from one. Both wrong answers to "whose credentials" are still the common ones.
Agents that share blind spots cannot catch each other's errors.
Three agents vote; majority wins. The green line is the textbook promise, computed exactly from the binomial: if errors are independent, three 80% agents vote their way to 90%. The red line is the same vote when the agents are correlated — when they err, they tend to err together, computed from a shared-cause model whose correlation you control. Agents built on the same foundation models, trained on the same data, prompted the same way, sit at the high-correlation end. There the ensemble gain rounds to zero: three agents confidently agree on the same wrong answer, and the vote launders one model's blind spot into consensus.
Further reading
- Wu et al. (2023), AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation — the framework, and a fair look at its patterns.
- Park et al. (2023), Generative Agents — the emergent-behaviour result that started much of the enthusiasm.
- Cemri et al. (2025), Why Do Multi-Agent LLM Systems Fail? — a taxonomy of failures observed in practice. Read this one first.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Adding agents to fix quality problems. If one agent is unreliable, three unreliable agents produce unreliable output more expensively.
- Assuming agents catch each other's errors. They share training data and share blind spots, so they often agree confidently and wrongly.
- Shipping without hard step and budget limits. Two agents can loop politely and indefinitely, and the bill arrives either way.
At a glance
Often compared with
Where this sits
A destination. 30 concepts lead here, and nothing in the corpus depends on it.
Computed from the prerequisite graph, not assigned. How this works