This is the long version of Building a Team of AI Agents — Explained Simply. Same starting point — multi-agent systems beat super-agent monoliths in production — but worked out at the level of architectural detail an engineer actually needs to ship one. Roles fleshed into a taxonomy. Memory architecture per role. The two relevant protocols (MCP and A2A) developed. Four orchestration patterns with the decision tree. Failure modes that are unique to multi-agent systems. The cost model. The evaluation discipline that distinguishes “the team converged on an answer” from “the team converged on a confident wrong answer.” Governance.

The natural companions are the Scaling Agentic AI / Production Playbook series for the broader production context, the API · CLI · MCP · ADK post for the protocol-layer treatment, and the Agent Frameworks monograph for the framework-by-framework comparison.

What it covers

Fourteen sections, about twenty-eight minutes of careful reading.

§ 1 — The case for a team. The cascading-error argument, fleshed out. The math. The honest comparison to the single-agent baseline.

§ 2 — The four canonical roles. Planner, Doer, Tool-user, Learner — fleshed into a working taxonomy. The responsibility of each. The interface each one exposes.

§ 3 — Memory architecture per role. Why each role needs a different kind of memory. The planner needs goals and constraints. The doer needs immediate context. The tool-user needs tool documentation. The learner needs the history of what worked and didn’t. Cross-reference to Knowledge & Memory in LLMs.

§ 4 — Inter-agent protocols. MCP for agent-to-tool communication. A2A for agent-to-agent. Why standardized protocols matter once you have more than two agents.

§ 5 — Orchestration patterns: sequential pipeline. The simplest. A linear handoff. When this works, when it doesn’t.

§ 6 — Orchestration patterns: supervisor-worker. The most common in production. One agent decides what to do; specialist workers do it. The handoff discipline.

§ 7 — Orchestration patterns: parallel fan-out. When you can decompose a task into independent sub-tasks and run them in parallel. The merge-step risk.

§ 8 — Orchestration patterns: hierarchical. Teams of teams. When this is the right shape and when it’s over-engineering.

§ 9 — Failure modes unique to multi-agent. Loops between agents (A asks B asks A). Stale information (A’s plan is based on B’s old output). Confidence amplification (the team agrees on a wrong answer because each agent saw the others’ agreement). The fixes.

§ 10 — Cost model. Token cost per agent. Why a team of small agents can be cheaper than one large agent, and the cases where it isn’t.

§ 11 — Evaluation for multi-agent systems. Why eval is harder when no single agent owns the final output. The per-role metric versus the system-level metric. The eval set that catches confidence amplification.

§ 12 — Governance & oversight. The human-in-the-loop placement. The audit log. The kill switch. The minimum disciplines that keep a team of agents from becoming a team of uncontrolled actors.

§ 13 — A worked example. A real multi-agent system traced end to end. A simple software-development workflow — planner agent decomposes a feature, doer agents implement, tester agent verifies, reviewer agent gates the merge. Real code, real failure modes, real fixes.

§ 14 — Further reading. The papers and write-ups that have aged well. The framework documentation that’s worth reading. The handful of post-mortems from teams who have shipped multi-agent systems and written honestly about it.

Read it

Open the monograph →

The monograph lives at its own URL with a warm-paper layout — terracotta-rust accent, deep teal cross-references, ochre highlights. Fourteen sections, four role chips color-coded throughout.


← Back to Autonomy