What is a multi-agent system in AI?
A multi-agent system is an AI application in which two or more language model agents, each with its own instructions, tools or context, split one task and coordinate through tool calls, handoffs, shared state or a protocol.
The term is older than large language models, but in 2026 it almost always means several LLM agents working on one job. The agents may be a lead and its subagents, a set of peers passing a conversation between them, or independent services that find and call each other over a protocol such as A2A.
All three model labs say to start with one agent and split only for a reason. OpenAI's practical guide says to maximise a single agent first. Google's Cloud Architecture Center lists the single agent as the starting point. Anthropic's October 2026 guidance goes further and recommends folding many subagent designs back into one loop.
Common designs
Manager or orchestrator
One agent keeps control and calls others as tools or spawns workers for subtasks.
Decentralised handoffs
Agents pass control of the conversation to one another as peers.
Sequential pipeline
Agents run in a fixed order, each working on the previous agent's output.
Parallel fan-out
Several agents work at once and a final step gathers their results.
Group chat and planners
Microsoft Agent Framework 1.0 (April 2026) ships sequential, concurrent, handoff, group chat and Magentic-One orchestrations as stable patterns.
What the evidence says
Google Research's controlled study of 180 configurations (January 2026) compared a single agent with independent, centralised, decentralised and hybrid multi-agent designs. Central coordination improved a parallelisable finance task by 80.9% over a single agent, but on a sequential planning task every multi-agent variant cut performance by 39% to 70%. A model built from task properties picked the right architecture for 87% of unseen configurations.
Cost is the other constant. Anthropic's write-up of its multi-agent research system (June 2025) measured multi-agent runs at about 15 times the tokens of a chat, so the design pays off only on valuable, breadth-first work.
How multi-agent systems fail
"Why Do Multi-Agent LLM Systems Fail?" by Cemri and colleagues (March 2025) annotated more than 1,600 traces from 7 popular frameworks and sorted the failures into 14 modes in three groups: system design issues, misalignment between agents, and weak task verification. In practice that means unclear roles, agents ignoring or repeating each other's work, and nobody checking whether the final result is right.
Google's study adds a structural warning: independent agents with no cross-checking amplified errors 17.2 times, against 4.4 times with central coordination.
How Motiontic's engineers apply it
Motiontic's engineers treat a multi-agent design as something to earn with evals. The default is one agent with good tools and skills; a second agent is added only for context isolation, true parallelism or conflicting instructions, with a central coordinator, scoped context for each agent and an end-to-end check on the final output.
Multi-agent coordination designs compared
| Centralised (manager or orchestrator) | Decentralised (peer handoffs) | Independent (no cross-checking) | |
|---|---|---|---|
| Who holds control | One lead agent | Whichever agent has the conversation | Each agent separately |
| How errors spread | The lead can catch and reconcile them (4.4 times amplification in Google's study) | Each agent passes on what it received | Unchecked (17.2 times amplification in Google's study) |
| Fits | Parallel subtasks with one owner of the answer | Specialists that each own a branch of a conversation | Rarely a good fit for one shared task |
| Main risk | An overloaded lead and high token cost | Lost context and checks that do not follow the conversation | Contradictory results nobody reconciles |
Common questions
Is a multi-agent system better than a single agent?
Only for some tasks. Google Research found big gains on parallelisable work and losses of 39% to 70% on sequential reasoning. Most teams get further by improving one agent's tools and context first.
What is the difference between a multi-agent system and an agentic workflow?
An agentic workflow fixes the path in code and may involve only one model. A multi-agent system has several agents, and at least some of the coordination between them is decided by models at run time.
How do agents in different systems talk to each other?
Through a protocol such as A2A, which lets agents advertise their capabilities and accept tasks over HTTP. A2A reached version 1.0 in March 2026 and joined the Linux Foundation's Agentic AI Foundation in August 2026.
Why do multi-agent systems fail?
Research by Cemri and colleagues groups the causes into poor system design, misalignment between agents, and weak verification of the final result. Lost context between agents and unclear ownership of the answer are the most common practical forms.
Related
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
Orchestrator-worker pattern
A lead agent plans, workers run in parallel with their own context, the lead merges.
Agent orchestration
The logic that decides what runs next, with what context, and who combines the results.
A2A protocol
The open protocol for agents talking to other agents, now under the Linux Foundation.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Google Research, Towards a science of scaling agent systems (2026-01-28)
- Anthropic, Multi-agent research system engineering post (2025-06-13)
- Cemri et al., Why Do Multi-Agent LLM Systems Fail? (arXiv, 2025-03-17)
- Google Developers Blog, Developer's guide to multi-agent patterns in ADK (2025-12-16)
- Microsoft, Microsoft Agent Framework Version 1.0 (2026-04-03)
- OpenAI, A practical guide to building agents (PDF, April 2025)
- A2A blog, Joining the Agentic AI Foundation (2026-08-27)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.