Single Agent or Multi-Agent? Start With the Simplest Design

By Mudabbir Hasan, Co-founder · Published 2026-10-06 · Last updated 2026-10-06

Start with the simplest design that works. Anthropic advises finding the simplest solution possible, often a single model call with retrieval, before a workflow or an agent. Multi-agent systems fit broad tasks that can be split into independent parts, but Anthropic reports they use about 15 times the tokens of a chat, so the task's value must justify the cost. Test on about 20 real cases before you scale.

What is the difference between a workflow and an agent?

Anthropic's engineering guide draws a line that most product descriptions blur. Workflows are systems where language models and tools are orchestrated through predefined code paths. Agents are systems where the model dynamically directs its own process and tool use. Both are agentic systems, but they behave differently: a workflow is predictable, and an agent decides its own steps. The guide was published in December 2024, so the tools it mentions have moved on, but the distinction still holds.

Why start with the simplest thing?

The guide advises finding the simplest solution possible and increasing complexity only when needed, which may mean not building an agentic system at all. For many applications a single model call with retrieval and examples is enough. It also suggests calling the model APIs directly before adopting a framework, because frameworks can hide prompts and responses and make debugging harder.

Climb only when needed

1
One model call with retrieval
Often enough.
2
A fixed workflow
Chain, route or parallelise: predictable steps.
3
A single agent
When the steps cannot be predicted.
4
Multiple agents
Breadth-first work that splits into independent parts.

Source: Anthropic.

Which patterns sit between one call and a full agent?

Prompt chaining: split a task into fixed sequential steps with optional checks between them. It trades speed for accuracy.
Routing: classify the input and send it to a specialised follow-up, such as sending easy questions to a smaller model.
Parallelisation: run independent parts at the same time and combine the results, or run the same task several times and compare.
Orchestrator-workers: a central model breaks the task down and delegates, for cases where the subtasks cannot be predefined.
Evaluator-optimiser: one model generates and another critiques in a loop, when you have clear criteria and iteration measurably helps.

When do multiple agents actually help?

Anthropic's account of its multi-agent research system is specific. Multi-agent systems excelled on breadth-first queries that pursue several independent directions at once, and on tasks with heavy parallelisation, information that exceeds one context window, and many complex tools. In its internal research evaluation, a lead agent with sub-agents beat a single agent by 90.2 percent. It is just as clear about where they do not help: tasks where agents must share one context or have many dependencies between them, and it gives coding as an example with fewer truly parallel parts.

What does it cost?

Anthropic reports that agents use roughly four times the tokens of a chat, and multi-agent systems roughly fifteen times. Its conclusion is that the value of the task has to justify that expense. The simple guide makes the same point about trade-offs: agentic systems often trade latency and cost for better task performance, and autonomy brings the potential for compounding errors, which it mitigates with testing in sandboxes, guardrails and stopping conditions such as a maximum number of iterations.

Anthropic's figures

90.2%
lead + sub-agents vs a single agent
Its internal research eval
≈ 4×
tokens: agent vs chat
≈ 15×
tokens: multi-agent vs chat
≈ 20
test cases to start with

What goes wrong in practice?

Over-delegation. Early versions spawned 50 sub-agents for simple queries, so effort has to be scaled to the difficulty of the task.
Vague instructions. Sub-agents given loose briefs duplicated work, so each needs an objective, an output format, tool guidance and clear boundaries.
Blocking. When work runs synchronously, the whole system can wait on one slow sub-agent.
Unpredictable behaviour. Small changes to the lead agent's instructions can change how the sub-agents behave.

Failure modes to plan for

Over-delegation: 50 sub-agents for a simple query
Vague briefs, so duplicated work
Blocking on one slow sub-agent
Small prompt changes shifting behaviour
Errors compounding in long runs

How do you test before you build?

Anthropic's advice is to start evaluating immediately with small samples of about 20 queries, because large effects show up with few cases. Judge outcomes rather than fixed paths, since agents can take different valid routes to the same result. A single model call grading against a rubric matched human judgement best in its tests, and human testing still found problems automated checks missed. Because errors compound in long-running agents, it also recommends checkpoints so a run resumes where it failed.

What does this look like in our work?

Our AI sales intelligence case divides the work by stage, with one agent for researching a prospect, others for identifying decision-makers, scoring leads, drafting outreach and monitoring replies, and handoffs between them. Our support case uses specialist agents for classification, knowledge retrieval, status lookups and internal operations, with escalation to a human when a case exceeds what an agent should resolve alone. Both engagements are confidential and the case pages describe what was built, not the orchestration pattern, so we do not claim which of the patterns above they use.

See the case study: AI sales intelligence platform

See the case study: support and operations agents

What should you do before you commit?

Try a single call with retrieval first and measure the result on about 20 real cases.
Move to a fixed workflow if the steps are known. Use an agent only if the number of steps cannot be predicted.
Add more agents only for work that splits into independent parts, and estimate the cost against the value of the task.
Decide the stopping conditions and the human handoff before you launch.

Read: When should an AI agent escalate to a human?

Custom AI and machine learning development