What is a Multi-Agent System?

How multi-agent systems work

Multi-agent systems split work across agents in one of two shapes. In the orchestrator/worker pattern, a lead agent decomposes the task, delegates scoped sub-tasks to workers, and synthesizes their results. In peer patterns, specialists hand work along a pipeline: a researcher passes to a synthesizer, which passes to a writer. Either way, agents coordinate through handoffs, and every handoff moves context from one agent’s window into another’s.

That transfer is where the architecture succeeds or fails. A UC Berkeley study of more than 1,600 traces across seven frameworks found failure rates as high as 86.7%, and most of the 14 failure modes it catalogs are inter-agent problems: one agent’s state contaminating another’s reasoning (bleed), full histories cascading between agents (explosion), and agents acting on superseded state (drift).

Why multi-agent systems matter

Splitting work lets each agent operate in a smaller, cleaner context than one agent juggling everything, which directly counters context rot. Done well, the gains are large: Anthropic’s research system, where each sub-agent works in its own window and returns only a 1,000 to 2,000 token summary, outperformed single-agent Claude Opus 4 by 90.2% on internal evaluations.

The catch is that handoffs are lossy. Causal reasoning, implicit constraints, uncertainty signals, temporal ordering, and rejected approaches all degrade predictably at handoff boundaries, and the loss compounds across hops. This is why the first agent in a pipeline is reliable and the fifth is not, even when each agent works correctly. Structured handoff schemas, not bigger models, are the fix.

The pattern is now moving into the models themselves. GPT-5.6’s Ultra mode decomposes a task into parallel subagents, each with its own context window, baked directly into the weights. That automates isolation but seals the handoff boundary where you can no longer inspect it.

Common misconceptions about multi-agent systems

• “More agents means better results.” Each added agent adds a handoff, and each handoff loses signal. If a single agent with good context engineering can do the work, the multi-agent setup is usually overhead.
• “The failing agent is the problem.” Usually the agent is fine and what it received is not. Diagnose the boundary before swapping the model.
• “Agents should share everything.” Shared state is how bleed and explosion happen. Production frameworks from Google, Anthropic, and LangChain all converged on passing compressed results, not raw context.
• “Isolation solves coordination.” Isolation fixes accumulation, not handoff. Sub-agents that never see each other’s decisions can make locally sensible, globally conflicting choices.

Multi-agent systems and Wire

Wire gives each agent in a multi-agent system its own scoped data boundary. Every container is an isolated, permissioned environment with its own MCP tools, so a sub-agent connected to one container cannot read into a sibling’s. Handoffs become explicit and inspectable: one agent writes results with wire_write, and the next retrieves exactly what it needs with wire_search or wire_navigate instead of inheriting a full history.