Production agentic systems use a coordinator that decomposes work and delegates to focused subagents, passing explicit context, with deterministic guardrails for high-stakes actions and human checkpoints for irreversible ones. These principles form the core of the Claude Certified Architect exam.
Hub-and-spoke orchestration
The dominant multi-agent pattern is hub-and-spoke: a coordinator agent in the middle, specialist subagents around it. All communication flows through the coordinator, which:
- decomposes the task and decides which subagents to call,
- passes each one the context it needs,
- spawns independent subagents in parallel to cut latency,
- aggregates results, spots gaps and re-delegates (iterative refinement).
Pipelines vs dynamic orchestration
- Fixed sequential pipeline (prompt chaining): same steps every time: extract → score → check → summarise. Predictable and auditable.
- Dynamic orchestration: the coordinator chooses agents at runtime for varied, open-ended requests.
Choose the simplest pattern that fits. Many business automations are pipelines, and that's a good thing.
Context and provenance
- Subagents are "amnesiac freelancers": include every fact they need in their prompt.
- Pass findings as structured data with sources (URL, document, date) so attribution survives synthesis.
- Keep a persistent case facts block (amounts, IDs, dates) that is never summarised away.
- When sources conflict, report both with attribution rather than silently picking one.
Guardrails and humans in the loop
- Deterministic over probabilistic when stakes are high: a refund limit belongs in a hook or prerequisite gate, not just a prompt.
- Human checkpoints for irreversible actions: deleting records, sending external emails, large payments.
- Escalate on the right triggers: the customer explicitly asks for a human, the request falls outside policy, or the agent can't make progress, not mere frustration or self-reported confidence.
- Audit everything: log tool calls and decisions for review and compliance.
- Measure by segment: 97% overall accuracy can hide a 40% error rate on one document type.
Check yourself
Three quick questions to check it stuck.
Ask the tutor
Stuck on something in this lesson? Ask the AI tutor. It answers from this lesson and points you to others that help. Answers are written by AI and can be wrong. Do not include personal details, passwords or company secrets.
Frequently asked
What is hub-and-spoke multi-agent architecture?
A coordinator agent sits at the centre and delegates tasks to specialist subagents. All communication routes through the coordinator, which decomposes work, passes context, aggregates results and handles errors.
When should an AI agent escalate to a human?
When the customer explicitly asks for a human, when the request falls outside documented policy, or when the agent cannot make meaningful progress. Sentiment alone and self-reported confidence are unreliable triggers.
How do I enforce business rules in an AI agent?
Use deterministic mechanisms such as hooks, tool interception or prerequisite gates for high-stakes rules like refund limits or identity checks. Prompts alone have a non-zero failure rate.