Claude agent architecture that survives real traffic

A Claude agent is a model plus a loop: observe, choose a tool or a reply, act, and stop. Production architecture is the set of limits around that loop so it cannot wander, overspend or write the wrong record.

Demos hide the loop. They run three happy-path steps on a staging tenant. Traffic introduces timeouts, partial payloads, ambiguous user intent and tools that return empty. The architecture has to assume those are normal.

The pieces that actually matter

Skip any one of those and you will relive why Claude pilots stall.

Single agent first

Start with one agent and a small tool set. Multi-agent Claude systems earn their complexity when you have a real split: research versus mutate, or a specialist that owns one system of record. Orchestration without evals is just more surface to debug.

Stop conditions are product decisions

"Keep going until it looks done" is not a stop condition. Define:

• Maximum tool calls per task
• Maximum wall-clock time
• When to ask a clarifying question
• When to hand to a human

Those rules belong in code, not in a paragraph of the system prompt. Prompts describe policy. Code enforces it.

If you need this loop built on your stack, that is the Production Agent Build: prototype by week two, live traffic by week six, evals in CI from the start.