Companion guide · 1 min · Reviewed Aug 2026
Agent Architecture
Single-agent tool loops, planners, memory stores, orchestrators, and multi-agent graphs are different shapes. Pick the smallest shape that makes the job inspectable. Read this alongside Agent Fundamentals and Agentic Workflows.
- Patterns
- State
- Verification
What you learn
How to choose a shape that stays inspectable as you add tools and roles.
Start with one agent
A single agent with a few tools is the correct default. Add roles when you have typed artifacts to pass, not when you want a crowd.
Visual
Single Agent with Tool Loop
The simplest architecture: one agent with a model and tools in a loop. Start here and only add complexity when you have typed artifacts to pass.
Make state visible
Memory, plans, and tickets should be data. If the only state is an endless chat, you cannot evaluate the system.
{
"ticket": "INC-4412",
"plan": ["reproduce", "patch", "test"],
"last_test": { "failed": 2 },
"stop": null
}Visual
Visible Agent State
Working state, memory, and plan artifacts are structured data — not buried in a chat transcript. This makes the system evaluable and debuggable.
Verification is a node
A check — tests, claim labels, policy — belongs in the architecture diagram. Hope is not a node kind.
FAQ
What is agent architecture?
Agent architecture is the arrangement of the model, tools, memory, and control flow that turns a task into observable steps and an output.