Books

I started writing Agentic Engineering because AI agents seem like one of the most consequential shifts in how software can be built and used. Their potential is that AI stops being a single-answer interface and becomes a coordinated system for long-running work: carrying context, using tools, gathering evidence, recovering from interruptions and helping people move through tasks that were previously too fragmented for one model call.

What interests me most is the engineering gap between that potential and dependable systems. The book grew out of my own work building ML systems, thinking about production reliability and watching agentic systems become powerful enough to need a clearer engineering language. I wanted to write the guide I would want beside me when turning an impressive agent demo into something genuinely useful.

The manuscript is a best-practices guide for building agentic systems that are capable, inspectable and dependable while still avoiding the real risks of opaque behavior, weak boundaries and unnecessary autonomy.

Book manuscript in progress

Agentic Engineering

Engineering Governable AI Systems for Operational Use

Agentic Engineering is about designing AI agents and agentic workflows that can plan, use tools, preserve evidence, recover from interruptions and support meaningful work in real operational settings.

The motivation is simple: agentic systems can turn AI from a passive interface into an active collaborator for research, operations and creative technical work. To make that promise real, teams need best practices for permissions, evidence, review, escalation and failure handling in addition to impressive demonstrations.

The central argument is that useful agency is an operating-model problem before it is a prompting problem. Autonomy becomes valuable when the system has a clear control surface, a visible proof surface and a workflow that lets people confidently supervise higher-leverage work.

The book is for engineers, product builders, founders and technical leaders who want to build agentic systems with ambition and discipline: systems that can take on larger tasks, use tools responsibly and improve the quality and speed of real work.

Main Questions

  • When is an agent the right abstraction compared with a conventional workflow or bounded model call?
  • Which decisions should remain explicit in the system design and which may be delegated to a model?
  • What artifacts demonstrate that the system performed useful work?
  • Where should human review, approval and escalation sit?
  • How should teams think about memory, tools, handoffs, observability, rollback and evaluation?

Core Themes

  • When agents are the right abstraction
  • Tool use and workflow design
  • Human supervision and review
  • Evidence, traces and evaluation
  • Reliability in production settings
  • Memory and context design
  • Patterns for responsible autonomy

Selected Book Concepts

  • Control surface
  • Proof surface
  • Context engineering
  • Tool authorization
  • Trace-based evaluation
  • Review gates
  • State and memory
  • Escalation paths
  • Failure recovery