Core agents
What is Agent Architecture?
Agent architecture is the arrangement of model, tools, memory, and control flow.
Architecture is independent of any one vendor. A tool loop, a planner, and a multi-agent graph are different architectures for different jobs.
Visual
Architecture spectrum
Progressively complex agent architectures from simple to orchestrated.
Why it matters
Architecture determines failure modes. A single-agent loop fails when the task outgrows one tool set. A multi-agent graph fails when handoffs drop state. Choosing the right architecture for the job prevents entire classes of errors.
Architecture is independent of vendor — the same patterns apply whether you use OpenAI, Anthropic, or open-source models.
Common patterns
The most common architectures are: single-agent with a tool loop, planning agent with explicit plan artifacts, memory agent with retrieval, and multi-agent systems with orchestrated handoffs. Each pattern adds capability at the cost of complexity.
The right choice depends on task scope, not on how impressive the architecture sounds. Most tasks need a single agent with two or three tools.
Key takeaways
- 1Start with the simplest architecture that can finish your task.
- 2Each architectural upgrade adds capability and failure surface.
- 3Architecture patterns are vendor-independent.
Common mistakes
- ✕Building a multi-agent system when a single-agent loop would work.
- ✕Choosing architecture based on impressiveness rather than task requirements.
Related terms
Concept neighborhood
Terms linked from Agent Architecture in the glossary graph.
- Agent Architecture
- Single Agent
- Multi-Agent System
- Orchestration