Architecting Agentic Systems Book
Engineering Dependable AI Agents
A technical reference for senior engineers and architects. It offers a formal pattern language for agentic systems, plus the architectural discipline that turns those patterns into reliable production systems: bounded autonomy, governance as architecture, memory and state, the ingestion pipeline, control and coordination, the skills layer, failure-mode taxonomy, testing and trace discipline, and the harness. The full text is free to read online.
Blog / Field Notes
Periodic commentary on what is moving in agentic systems. A companion to Architecting Agentic Systems, time-bound where the book is stable.
Who can turn off an agent's approval checks?
August 28, 2026
An August discussion of the Hermes attack focused on unattended execution. IBM's production account put access decisions at the backend. Together they sharpen the question of whose authority an agent uses.
The tool call timed out. Did the action happen?
July 18, 2026
July discussions about agent frameworks and durable loops raised a specific recovery problem: a timed-out tool call may already have changed the world. What makes the next attempt safe?
July 5, 2026
A practitioner argues for simpler automation; a June study examines the work of supervising agents. Both sharpen the workflow decision: which choices need a model, and what will reviewing them cost?