Multi-agent systems are technologically exciting. When implementing them for a client, however, it often turns out that a simpler approach would have handled 90% of requirements at 20% of the cost. Here's a decision map for when to invest in multi-agent orchestration and when not to.
Single agent vs. multi-agent — where the line is
**Single agent** = one LLM (Claude, GPT, Llama) with a few tools. Example: a chatbot that searches documents, generates an answer, optionally calls an API.
**Multi-agent** = orchestration of multiple agents, each with its specialization, and a meta-agent that decides which one continues the task. Example: one agent plans, another codes, a third tests, a fourth reviews.
A single agent is always cheaper — one model, one flow, one audit log. Multi-agent is always more complex — 3–8× more model calls, 5× more latency, 10× more potential failures.
But multi-agent delivers things a single agent cannot: long multi-step tasks, parallel work, oversight over decisions.
When a single agent is enough
1. The use case has a clear, predictable structure.
Example: "answer a client email based on the context in CRM". Few steps, simple decision-making, response quality verified by a human reviewer.
A single agent with prompt templates and 2–3 tools (CRM search, context memory, formatting) handles 90% of the volume.
2. The team doesn't have capacity for an orchestration layer.
Multi-agent requires: a workflow engine (LangGraph, Temporal, Inngest), agent state management, error handling, retry logic, observability. Without a dedicated AI engineer, the system stops at the first degenerate step.
If your team is fewer than 3 engineers, single agent is the realistic level.
3. The use case has an acceptable failure rate of 5–10%.
A single agent fails gracefully — the answer is worse, but the UI doesn't break. Multi-agent fails compoundly — one failed agent in the chain breaks all subsequent ones. Without sophisticated error handling, multi-agent delivers worse than a single one.
When multi-agent makes sense
1. The use case has long tasks (10+ steps) with different capabilities.
Example: an agent firm implementing a feature: a planner analyzes the requirement, a developer writes code, a tester writes tests, a reviewer checks the PR. A single agent cannot handle this with reasonable quality — context dilutes, the model loses the trail.
2. Acceptable failure rate is < 1%, or the task is business-critical.
Multi-agent allows human-in-the-loop gates between agents. Planner proposes a plan → human approves → developer implements → reviewer checks → human approves the merge. A single agent doesn't model this.
3. Parallel work brings measurable benefit.
Multi-agent can branch a task (one agent investigates 5 sources in parallel, one agent generates 3 versions of text). A single agent must run sequentially — slower, but simpler to debug.
4. Audit trail is a regulatory requirement.
In regulated sectors (law, healthcare) a per-step audit log is mandatory. Multi-agent naturally generates an audit log per step — agent A did X, agent B reviewed X, agent C approved. Single agent log is "the answer is…" without details in between.
The cost nobody mentions
A multi-agent system is not only more expensive on inference (5× more LLM calls). Costs that appear in months 2–6:
- **Prompt drift**: when one agent is changed, all subsequent ones need to be retested. Without systematic evals, this means hours of manual testing on every change.
- **State explosion**: agent state in complex workflows has 50+ variables. Without good state management, the system becomes unmaintainable.
- **Observability**: without detailed logging you can't debug why the system produces worse answers. An observability stack (LangSmith, Helicone, Trackio, custom) costs €200–1,500 per month.
- **Human-in-the-loop UI**: if you want human gates, you need a UI that enables them. A custom dashboard for approval flow costs €5,000–15,000.
Practical approach
**Step 1**: Start with a single agent. It covers 60–80% of the use case. It identifies which steps are done poorly.
**Step 2**: For those specific steps, add specialized agents. Multi-agent "minimum" — 2 agents + one orchestrator.
**Step 3**: Add more agents only if measurable benefit > cost of additional complexity.
Most multi-agent systems in 2026 in production are 2–4 agents, not 10+. Team size is not a measure of complexity — it is a measure of what you haven't optimized.
Conclusion
Multi-agent is not "better" than single agent. It is a more expensive investment for capabilities that some use cases need and others don't.
- If your use case is predictable, short, low-risk → single agent.
- If long, multi-stage, business-critical → multi-agent is worth the investment.
- If you're in between — start with a single agent and gradually add specialized agents.
---
*We implement both single-agent and multi-agent systems. We make the choice through a business-case audit, not through technological preference. We can walk through an example of a specific use case on a 30-minute call.*