Skip to content

Ch 16: Multi-Agent Systems Architecture - Advanced

Track: Advanced | Try code in Playground | Back to chapter overview

Read online or run locally

You can read this content here on the web. To run the code interactively, either use the Playground or clone the repo and open chapters/chapter-16-multi-agent-systems-architecture/notebooks/03_orchestration_and_evaluation.ipynb in Jupyter.


Chapter 16: Multi-Agent Systems — Notebook 03 (Orchestration & Evaluation)

Finally we wire agents into workflows — supervisor, pipeline, debate — and measure them on success, cost, and latency, while guarding against deadlock and runaway loops.

What you'll learn

Topic Section
Pipeline, supervisor, and debate topologies §1
Step budgets and loop guards §2
Cost and latency accounting §3
Evaluating a multi-agent system §4

Time estimate: 2.5 hours


Key concepts

  • Topology — how control flows: linear (pipeline), star (supervisor), or adversarial (debate).
  • Step budget — a hard cap that stops infinite delegation and runaway cost.
  • Cost accounting — every agent call is money and latency; measure it.
  • Evaluation — judge the system on task success, cost, and latency, not single outputs.

Topology determines how control and cost flow through a multi-agent system. Step budgets prevent the classic runaway-loop failure, and honest cost/latency accounting is what separates a demo from a deployable system.

Run the full notebook in the chapter folder for code and outputs.


Generated by Berta AI