Guides · Series · 6 parts
Federated AI Agents
Every call that crosses from one owner to another is a seam, and a seam has five properties — contract, authority, state, failure, and accountability — that are free inside one deployment and have to be built across it.
The book
Federated AI Agents
LangGraph and A2A for Agents You Don't Own: Build, Secure, and Operate the Seams Between Teams and Companies
A production guide to the A2A (Agent2Agent) protocol and LangGraph for AI agents that different teams and companies own.
Agents you do not own
Your LangGraph agent now has to call an agent somebody else owns: another team’s service, shipped on that team’s schedule, or another company’s agent behind its own server. Inside one deployment you get a great deal for free. One checkpointer sees every write, one credential covers every call, and one stack trace shows where something broke. None of that crosses an owner line.
Multi-agent frameworks assume you own every agent. Your partner releases on its own schedule, keeps its own state, answers to its own auditors, and sends you its own invoice. This series shows what to build on your side so that you can call it anyway.
Every part is worked on one running example, Meridian Mutual, a motor insurer whose claims agents are split across teams and a repair-estimate partner at another company.
Which part solves my problem
The series reads in order, but if something is already broken, start here.
- “I cannot tell whether a call should be a subgraph, an MCP tool, or an A2A seam.”Part 1 →
- “I need to call a remote A2A agent from inside a LangGraph graph.”Part 2 →
- “A partner’s behaviour changed while its agent card did not.”Part 3 →
- “A retry across the seam may have done the work twice.”Part 4 →
- “Work the partner already committed has to be undone.”Part 5 →
- “I have to test against a partner I cannot deploy.”Part 6 →
The series
0 of 6 published
Multi-Agent Is Not Federated: The Federation Test
When a call should be a subgraph, an MCP tool, or an A2A seam, and what each choice costs.
The A2A Call as a LangGraph Node
Wrapping a remote agent in a node, and what the checkpointer stops seeing when you do.
The Agent Card Is a Contract
Treating the card as a versioned contract you test against, not as discovery metadata.
Retries, Idempotency, and Reconciliation Across the Seam
Why an idempotency key protects you only if the partner honours it, and how to reconcile against the task ledger.
Compensation Across Owners
Undoing committed work at another company, where compensation is a request, not a rollback.
Testing Agents You Can’t Deploy
Stubs and contract tests for A2A partners you cannot run in CI.
Parts still marked in progress are covered in full by the book today — the chapter ranges above are exact. The articles are being written against those chapters.
The companion code
The companion repository builds the whole federation chapter by chapter in Python, and its test suite runs with no network and no accounts.
View the companion repository on GitHub →Other guide series
GenAI Architectures
There are eight architectures and seven boundaries between them. Every boundary is a yes-or-no question about the task — not a judgment call about the design.
Agentic AI with LangGraph
An eight-part guide to the part of agent engineering nobody demos: making the thing survive a restart, a concurrent write, a bad tool call, and a Tuesday deploy.
LangGraph Agent Evals
Every layer of an agent eval stack is a measuring instrument with its own error rate — the grader included — and none of them should gate a deploy until it has been scored against known faults.
Harness Engineering
AI systems do not fail at the model layer. They fail in the code around it — and that code is a designable layer with its own architecture.
RAG Engineering
Most RAG failures happen upstream of the LLM, accumulate invisibly, and only surface as an incident. This series names each one.
Claude Code Playbook
Claude Code has four layers — context, skills, subagents, hooks — and the variable that decides which one you need is enforceability, not capability.
AI Control Plane
A fleet of agents needs a layer that is not itself an agent — observability, policy, orchestration, versioning, cost, and audit, built as infrastructure.
Agentic AI in the Data Center
In physical infrastructure, an agent’s autonomy is capped by reversibility and jurisdiction — not by how good the model is.
Codebase to Architecture Doc
A three-pass methodology for compressing any codebase into validated diagrams and decisions — and then a LangGraph pipeline that runs it for you.
