Interest in multi-agent systems has exploded, and frameworks make it easy to wire up a planner, a researcher, a coder and a reviewer that talk to each other. It looks impressive. It is also frequently the wrong starting point.
Every extra agent adds latency, token cost, hand-off errors and another thing to debug. Start simple and add agents only when you can name the problem they solve.
The ladder of complexity
Climb one rung at a time.
- A single model call with a good prompt. Many "agent" problems are really this.
- A workflow: fixed steps in code, with a model call at some of them. Predictable and easy to test.
- A single agent with tools, choosing its own steps in a loop.
- Multiple agents, each with its own context and tools, coordinated by an orchestrator.
When multiple agents genuinely help
- Context isolation. A sub-task produces lots of noise (web research, log analysis) that would pollute the main agent's context. A sub-agent digests it and returns a summary.
- Parallel work. Independent sub-tasks can run at the same time, such as researching five competitors.
- Different permissions. A reader agent can access untrusted content with no write tools; a separate agent with write access only sees the reader's vetted summary. This is a useful defence against prompt injection.
- Specialised tool sets. Each agent gets a small, focused toolset instead of one agent choosing from fifty.
When they hurt
- Agents need to share lots of state, so every hand-off loses detail.
- The task is sequential and each step depends tightly on the previous one.
- You cannot yet evaluate the single-agent version. Adding agents multiplies what you cannot measure.
Design the hand-offs
Most multi-agent failures happen at the boundaries.
- Define a structured result for each sub-agent: what it found, confidence, sources, open questions.
- Give sub-agents a clear stopping condition and a budget of steps or tokens.
- Keep the orchestrator responsible for the final answer, so there is one place to check quality.
{
"task": "Summarise breaking changes in library X v5",
"result": ["Removed default export", "Node 18 no longer supported"],
"sources": ["CHANGELOG.md#v5.0.0"],
"confidence": "high",
"open_questions": []
}Measure before and after
Compare the multi-agent design against the single-agent baseline on the same eval set: success rate, cost per successful task and latency. If the gain is small, keep the simpler system.
Key takeaways
- Start with a single call or a fixed workflow, then climb only when needed.
- Multi-agent designs shine for context isolation, parallelism and permission separation.
- Structured hand-offs and step budgets prevent most coordination failures.
- Prove the gain with evals before accepting the extra cost.