Path · 7 steps
Agents in production, without a shared robot
- Steps
- 7
- Reading
- About 9 min
- Disciplines
- 7
New to the words? The glossary explains terms like RAG, grounding and eval harness in a paragraph each.
The problem
A supervisor and a handful of specialists worked in the demo. Then it needed real permissions, real identity, retries that do not double-charge, and an audit trail a reviewer would accept.
When you finish
You can explain why one gateway, per-user identity and an outbox are required for agents that write, and point at code for the orchestration.
- 1ArchitectureDesigning a Multi-Agent System for Real-World Use
- 2CodeMulti-Agent Orchestration (Python)
- 3DecisionLangGraph vs the OpenAI Agents SDK
- 4SecureYour agent should not be a superuser
- 5DeployA retried agent job is a second chance to send the same email
- 6ObserveWhat an AI gateway actually costs to run
- 7LessonsYour tool registry is an access control list wearing a different name
Architecture · step 1
Designing a Multi-Agent System for Real-World Use
The supervisor pattern and where humans belong. Every later step assumes it.
Code · step 2
Multi-Agent Orchestration (Python)
Run the pattern in thirty lines before adding any infrastructure.
Decision · step 3 · Comparison
LangGraph vs the OpenAI Agents SDK
Choose the orchestration runtime on control flow and tracing.
Secure · step 4 · Explainer
Your agent should not be a superuser
The identity mistake almost every first deployment makes, and the on-behalf-of fix.
Deploy · step 5 · Explainer
A retried agent job is a second chance to send the same email
Agents retry, so every write tool needs a key before the first retry happens.
Observe · step 6 · Article
What an AI gateway actually costs to run
The operational bill for the gateway, so it survives the budget review.
Lessons · step 7 · Explainer
Your tool registry is an access control list wearing a different name
The shape you end up with if you skip the steps above.