One agent is often easier to operate when the task shares one context and one authority boundary. Add specialists only when they provide a measurable benefit that justifies more handoffs, latency, cost, and failure states.
Agent count is an architectural choice
A single-agent workflow uses one model-controlled worker for the bounded task, with trusted tools and state outside it. A multi-agent workflow divides interpretation or planning among several workers. Neither arrangement removes the need for backend permissions and accurate outcome evidence.
For a proposed comparison, begin with the simplest design that meets the requirements. Identify the specific reason for splitting work: independent expertise, separable evidence retrieval, different access scopes, or parallel tasks whose results can be combined safely.
Do not create extra agents merely to rename stages in a linear prompt. If every specialist needs the entire history and can perform the same tools, the split may add coordination without meaningful separation.
Compare a fictional repair intake
The task collects contact details, validates service area, confirms the request, and saves it. A single agent with deterministic validation tools may handle this clearly. Splitting name collection, address interpretation, and confirmation into separate agents can create repeated questions and stale handoffs.
A different task compares independent technical documents under separate access scopes. Specialists may provide a useful division if their evidence contracts and final ownership are explicit. The benefit should appear in controlled outcome tests, not only in a more elaborate diagram.
| Comparison dimension | Evidence |
|---|---|
| Correct business outcome | Authoritative record assertions |
| Handoff failure | Missing or stale context cases |
| Latency | End-to-end measured task duration |
| Cost | Included model, tool, and recovery work |
| Operability | Staff ability to diagnose and recover |
Run both designs against the same fictional fixtures, including corrections, tool outages, and ambiguous evidence. Avoid comparing a polished multi-agent implementation with an intentionally weak baseline.
Keep future complexity reversible
A well-defined tool and state boundary can support later specialisation without changing the business contract. Start with clear interfaces and add agents when a demonstrated bottleneck or requirement warrants them.
A single agent can still become overloaded by long context or incompatible responsibilities. The decision is empirical and scope-specific. Report the tested conditions and avoid universal claims that one pattern is always cheaper or more reliable.
This architecture method is proposed. The voice platform case study describes related integration work without claiming comparative agent benchmarks. See specialist routing boundaries.
Single-agent FAQ
Does one agent mean no orchestration?
No. Tools, durable state, permissions, and recovery still need application control.
When should specialists be added?
When a clear separable requirement or measured benefit justifies the coordination cost.
What should determine the winner?
Use correct outcomes, critical defects, latency, total cost, and operational clarity under the same fixtures.