Agentic Workflows

Who this series is for

This series follows Building AI Agents. It is for engineers whose single agent has become too unpredictable to trust with a multi-step business process. The central idea is that most production agentic systems are mostly deterministic workflows with a few carefully placed model decisions, not fully autonomous agents.

What each chapter covers

• Workflow Graphs covers modelling a process as explicit states and transitions, so each step is testable, resumable after a failure, and observable.
• Multi-Agent Systems covers when splitting work across specialised agents genuinely helps, when it only multiplies cost and failure points, and how to coordinate agents that do need to cooperate.
• Model Context Protocol (MCP) covers exposing tools and data sources through a standard interface, so the same integrations work across different models and clients.

How to approach it

Pick a real process with at least one point where a human should approve something, and model it as a graph before writing any code. Most of the design value comes from deciding which steps should not involve a model at all. Allow roughly eight hours.