Multi-agent systems were the single most hyped AI architecture pattern of 2025. A year into serious production deployment, the data is messier than the keynote slides suggested, but the genuine wins are starting to come into focus.

Where multi-agent works

The clearest production wins are in domains with naturally decomposable workflows where the agents do not need to negotiate with each other very much. Customer service triage and routing. Document processing pipelines with specialist extraction agents per document type. Research and synthesis workflows where a planner agent dispatches to retrieval agents and a writer agent assembles.

These are essentially distributed task systems with an LLM controller. They work because each agent has a bounded task, the failure modes are well understood, and the coordination overhead is light. The teams shipping these are reporting the same kinds of accuracy and cost wins that single-agent systems delivered, with better latency on parallelisable workloads.

Where multi-agent struggles

Negotiation-heavy multi-agent setups are still mostly demoware. When two or more agents need to actually coordinate — debate a plan, refine outputs across multiple turns, agree on a path — the failure modes pile up. Token cost balloons. Latency gets unpredictable. Agents get stuck in loops. The benefits over a well-designed single-agent system with multiple tools are small or negative.

The pattern that has emerged: if you can express the workflow as “specialist sub-agents called by a controller,” it tends to work. If you need genuine multi-agent negotiation, you are in research territory still.

The cost story

Multi-agent systems have a worse cost story than the marketing acknowledges. Two to four agents per request, several tool calls per agent, sometimes a verifier agent on top — the token spend per task can be five to ten times a single-agent equivalent. For high-value workflows this is fine. For volume workflows it is a problem.

Teams shipping multi-agent in volume have invested heavily in agent-level caching, in deciding when not to call sub-agents, and in distillation from larger to smaller models within the agent stack. None of that is in the off-the-shelf frameworks.

The recommendation for 2026

Multi-agent is real, but use it where the workflow naturally decomposes. Treat the orchestration code as production software, not a notebook. Budget for the eval overhead, which is harder than single-agent eval.

The hype will fade. The architecture pattern will stay.