Agentic architecture

Agentic architecture: choose autonomy deliberately.

The best architecture is not the one with the most agents. It is the smallest system that can make the required decisions, use the necessary tools, and expose enough evidence for a person to trust and improve it.

Workflow or agent?

A workflow follows predefined paths through code. An agent chooses its own next step in response to what it observes. Workflows are easier to predict and test; agents are useful where the path cannot be fully enumerated in advance. In practice, reliable systems often combine a deterministic shell with tightly bounded agentic steps.

Five architecture decisions

Start with the smallest control loop

Use a deterministic workflow when the steps, transitions, and failure behaviour are known. Add agentic autonomy only where the system genuinely needs judgment over uncertain intermediate results.

Separate planning from execution

A plan is a proposal; execution changes state. Keeping them distinct makes review, permissions, retries, and audit trails much easier to reason about.

Give state a single owner

Specify whether a fact lives in application state, a tool response, a durable store, or the current conversation. Multiple implicit owners create contradictory decisions.

Bound every loop

Set a stopping condition, time or step budget, tool-call limits, and a clear escalation path. An agent should be able to stop safely as well as succeed.

Evaluate the whole trajectory

For an agent, a fluent final answer is not enough. Test tool choices, state changes, safety boundaries, recovery behaviour, and the evidence behind the outcome.

A design review checklist

  • What decision requires model judgment, and what can remain deterministic?
  • Which tool or data source is authoritative for each important fact?
  • What is the maximum cost, number of actions, and time an execution may consume?
  • What evidence would prove that the system made a useful and safe decision?
  • When should the workflow pause and return control to a person?

Primary sources to verify

Use the architecture review checklist

Turn these design principles into a concrete pre-ship review for goals, state ownership, tools, safety, limits, and evaluation.

Open the Claude system design checklist →

Test your architecture judgment.

Use original coordination scenarios to practise choosing ownership, limits, and human handoffs.

Practise multi-agent questions →