Single agent vs multi-agent: one loop first, split only for a reason
A single-agent system runs one model loop with its own tools and context, while a multi-agent system splits work across several agents that each hold their own context and pass results between them; Anthropic, OpenAI and Google all advise starting with one agent and splitting only for context isolation, parallel work or conflicting instructions.
Checked
Multi-agent designs look appealing on a whiteboard: a planner, a researcher, a writer, a reviewer. In production they cost more tokens, add handoffs where errors creep in, and are harder to debug. The labs' guidance in 2026 has moved toward fewer agents, not more.
Anthropic's newest guidance, from October 2026, says many enterprise agents work better as one loop with skills and tools loaded on demand. Google Research's controlled study shows where multi-agent helps and where it hurts.
What the evidence says
Anthropic, 2025
Its research system uses a lead agent that spawns parallel subagents, each with its own context. Multi-agent runs used about 15 times the tokens of a chat, so the pattern pays off only on high-value, highly parallel, breadth-first work. Coding was a poor fit.
Anthropic, October 2026
For many enterprise agents, one loop with progressively disclosed skills and tools does better than orchestrator-plus-subagent designs on quality, cost and latency. One customer saw roughly twice the speed. It warns against subagents shaped like an org chart.
Google Research, 2026
Across 180 configurations, multi-agent coordination helped a lot on parallelisable tasks but cut performance by 39 to 70% on sequential reasoning. Independent agents with no cross-checking amplified errors about 17 times; centralised coordination contained them.
OpenAI
Maximise a single agent first. Move to multiple agents only when logic or tool overlap gets too complex, using a manager agent that calls others as tools, or handoffs between peers.
When splitting is justified
Context isolation
A subtask would flood the main context, such as reading hundreds of documents to return a short answer.
True parallelism
Independent subtasks that can run at the same time, such as researching many companies at once.
Conflicting instructions
Two roles whose rules would contradict each other in one prompt, such as a generator and a strict reviewer.
Ownership boundaries
Agents built and run by different teams or vendors, where a protocol such as A2A connects them.
If the system does split
Use a central orchestrator rather than a free-for-all of peers, give each subagent only the context it needs, and have it return a condensed result. Anthropic advises migrating one subagent at a time and comparing evals at each step.
Motiontic's engineers start every build as one agent with a small tool set and skills, and add a subagent only when an eval shows a specific gain in accuracy, cost or speed that justifies the extra handoff.
A single-agent system compared with a multi-agent system
| Single agent | Multi-agent | |
|---|---|---|
| Context | One shared window, managed by the harness | Separate windows per agent |
| Token cost | Lower | About 15 times a chat in Anthropic's research system |
| Fits | Sequential tasks and tool-heavy work | Parallel, breadth-first work |
| Error behaviour | Errors stay in one trace | Up to about 17 times amplification without cross-checks (Google Research) |
| Debugging | One trace to read | Several traces plus the handoffs between them |
| Lab default | Start here (Anthropic, OpenAI, Google) | By exception, with a central orchestrator |
Common questions
Are multi-agent systems more accurate than single agents?
Only on some tasks. Google Research found multi-agent coordination helped on parallelisable tasks but cut performance by 39 to 70% on sequential reasoning, and Anthropic's 2026 guidance finds one loop often does better on quality, cost and latency.
How much more do multi-agent systems cost?
Anthropic reports its multi-agent research system used about 15 times the tokens of a chat. The extra cost is justified only when the task is high-value and can be split into parallel parts.
When does Anthropic recommend subagents?
For context isolation, parallel work and conflicting instructions. Its October 2026 guidance otherwise recommends folding subagents back into one loop with skills and tools loaded on demand.
What did Google's study of agent systems find?
Across 180 configurations, multi-agent setups helped on parallelisable tasks, hurt on sequential reasoning, and amplified errors about 17 times when agents worked independently without cross-checking. Centralised coordination contained errors most effectively.
How does Motiontic decide between one agent and several?
Every build starts as one agent with a small tool set and skills. A subagent is added only when evals on the client's real tasks show a clear gain that outweighs the extra cost and handoff.
Related
Workflow vs Agent
A path fixed in code, compared with a model that chooses its own steps.
MCP vs A2A
The protocol for tools and data, compared with the protocol for agents talking to agents.
Skills vs MCP
Connectivity to systems, compared with packaged expertise the agent loads on demand.
Agent harness
The loop, tools, context, permissions and state that sit around the model.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
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.