Core agents
What is Single Agent?
A single agent is one model-driven controller with tools and state for the whole task.
Single-agent loops are the default. Add more agents only when typed handoffs buy you something a bigger tool set cannot.
Why it matters
A single agent is the default architecture for 80%+ of agent tasks. It has one model, one set of tools, and one control flow. It is easier to debug, test, and monitor than any multi-agent alternative.
The temptation to add more agents is real but usually premature. A single agent that fails gracefully is better than a multi-agent system with lost handoffs.
When to upgrade
Move beyond single-agent when: the task requires distinct specialist roles that produce different artifact types, the tool set is too large for one context window, or you need parallel execution. If you just need more tools, add tools — don't add agents.
Key takeaways
- 1Single-agent is the default. Promote to multi-agent only when you need typed handoffs.
- 2More tools ≠ more agents. Add tools to a single agent first.
- 3One trace is easier to debug than many.
Common mistakes
- ✕Adding a second agent when adding a second tool would solve the problem.
- ✕Treating single-agent as "beginner" and multi-agent as "advanced" — they are different solutions for different problems.
Related terms
Concept neighborhood
Terms linked from Single Agent in the glossary graph.
- Single Agent
- Agent
- Tool Loop
- Multi-Agent System
FAQ
- Is a single agent less capable than a multi-agent system?
- Not inherently. A single agent with the right tools can outperform a poorly designed multi-agent system. Capability depends on architecture fit, not agent count.