The industry moved from single-prompt workflows to multi-agent orchestration fast enough that "how many agents" became a design decision before most teams had evidence for an answer. The honest position: more agents buys specialisation and costs you determinism, and most systems that fail in production fail on determinism.
The three shapes
- Single agent with tools. One model, one loop, a set of tools. Easiest to debug because there is one trace and one place a decision was made.
- Planner-executor. A capable model produces a plan; cheaper models execute the steps. This is the pattern behind most of the cost reductions people report, because the expensive model runs once rather than at every step.
- Orchestration graph. Multiple specialised agents with handoffs, often with a supervisor. Real capability gains on genuinely heterogeneous work, and a step change in how hard failure is to attribute.
Start with one agent
A single agent with well-described tools solves more than teams expect. The instinct to split into a "researcher", a "writer" and a "reviewer" usually encodes an org chart rather than a technical constraint, and each handoff is a place context is dropped and latency is added.
Split when you have a concrete reason you can name: genuinely different tool permissions, a step that needs a different model class, or a review step that must not share the generator's context.
The cost argument for planner-executor
The economics are the strongest case for splitting. Heterogeneous architectures — a frontier model for planning and orchestration, mid-tier for standard work, small models for high-frequency execution — are how teams keep agent costs survivable at volume. Reported reductions from the plan-and-execute pattern reach 90%, because the expensive reasoning happens once instead of on every tool call.
Route on task, not on prestige
The routing rule that works: send the ~80% of predictable, well-specified steps to the smallest model that passes your evals, and escalate only the genuinely ambiguous remainder. This requires you to have evals first — without them you are guessing which steps are safe to downgrade.
What multi-agent actually costs you
- Attribution. When a five-agent workflow produces a wrong answer, the failing step is not obvious from the output. You need trace-level evaluation to find it.
- Context loss. Every handoff is a summarisation, and every summarisation drops something. Failures often trace to information that existed two agents ago.
- Latency multiplication. Serial handoffs add up; users feel it.
- Non-determinism compounding. Each agent's variance multiplies rather than averages out.
A decision rule
Use one agent until you can state, in a sentence, what a second agent does that the first cannot. "Separation of concerns" is not that sentence — it is an aesthetic preference imported from code architecture, where components are deterministic. If the reason is different permissions, a different model class, or independent review, split. Otherwise the second agent is added surface area.
Frequently Asked Questions
Is multi-agent always more capable?
No. It is more capable on heterogeneous tasks that genuinely need different tools or model classes. On homogeneous tasks it usually adds latency and failure modes without adding capability.
How do I debug a multi-agent failure?
Trace-level evaluation: score each step — the plan, the tool calls, the handoffs — not just the final output. End-to-end scoring tells you it failed, not where.
Does the planner have to be the most expensive model?
It has to be the most capable at decomposition, which is usually but not always the most expensive. Test it — a mid-tier planner that passes your evals is a large saving.
References
- 7 Agentic AI Trends to Watch in 2026 — Machine Learning Mastery
- Top 15 Agentic AI Trends to Watch in 2026 — Firecrawl
About Jishu Labs
Jishu Labs is a software development company founded in 2016. We build custom software, AI/ML systems, and full-stack web and mobile applications for clients, and we make eight AI tools for software teams.