The Art of Agents

Building Agentic AI Systems That Think Before They Code

13 chapters mapping Sun Tzu's principles to modern agentic AI design. Each chapter teaches a concept, builds a piece of a running platform, and closes with a real field report. By the final chapter, you have a working multi-tenant experimentation platform.

Read online Get notified of updates

XIII Chapters

Chapter I

Laying Plans

Spec-driven design. The five constants of agentic systems: Contract, Context, Terrain, Model, and Protocol.

Chapter II

Waging War

The cost of agentic systems. When to build, when to buy, and when to walk away.

Chapter III

Attack by Stratagem

Defence is schema. Input validation, output parsing, tool-call whitelisting, scope constraints.

Chapter IV

Tactical Dispositions

Defence through architecture. Walls and moats that protect from adversarial users and agent improvisation.

Chapter V

Energy

Tool design. A well-designed tool multiplies capability by orders of magnitude. MCP as the standard protocol.

Chapter VI

Weak Points & Strong

Observability. Every agent invocation produces a trace. The undefended places are the ones you cannot see.

Chapter VII

Manoeuvring

Adaptive strategy. Graceful degradation, fallbacks, and the ordered retreat when things go wrong.

Chapter VIII

Variation in Tactics

Agent formations. Solo, pipeline, swarm, hierarchy — choose based on terrain, not fashion.

Chapter IX

The Army on the March

Deployment. Shadow mode, gated rollout, kill switches, rate limits, and rollback plans.

Chapter X

Terrain

Auth boundaries, data governance, organisational trust. Map the human terrain as carefully as the technical one.

Chapter XI

The Nine Situations

Nine failure modes. For each, the spec should define expected behaviour before the agent encounters it.

Chapter XII

Attack by Fire

LLMs are fire. Extraordinarily powerful in the right conditions, catastrophically wasteful in the wrong ones.

Chapter XIII

The Use of Spies

Feedback loops. Human feedback, automated evaluation, agent self-reflection, and continuous improvement.

Preview: Chapter I — Laying Plans

Sun Tzu opens with a warning: war is a matter of life and death, and must never be entered into lightly. The same applies to building agentic systems.

An agent released without a specification is a soldier sent to battle without orders — dangerous to friend and foe alike. Write the spec first. Define what the agent can do, what it cannot do, and what it should do when it encounters the unexpected.

The five constants of agentic design map directly from Sun Tzu: the Contract (the unambiguous agreement between builder, agent, and world), the Context (the data landscape, latency constraints, regulatory climate), the Terrain (APIs, databases, file systems, human interfaces), the Model (the reasoning engine, its capabilities, biases, and cost profile), and the Protocol (tool use, memory management, retry logic, guardrails).

You don't spec then build. You build to discover the spec.