For those who build it

The Craft

The practitioner's complete, concept-level playbook of AI today — novelties included.

An agent is a model plus a harness. Everything hard lives in the harness.

Agents vs Workflows

The foundational distinction — and the discipline of choosing.

The Five Patterns

Chaining, routing, parallelisation, orchestrator-workers, evaluator-optimiser.

Harness Engineering

Guides and sensors — the new senior discipline of the field.

Memory & Context

The four memory types and the four operations of context engineering.

Tool design principles and the standard interface for exposing them.

RAG & Its Evolutions

From naive top-k to agentic, graph, and self-correcting retrieval.

Multi-Agent, Handoffs & A2A

Topologies, transfer contracts, and the inter-agent wire protocol.

Mechanistic activation control and operational instruction-fidelity.

Code & Doc Indexing

Indexed vs runtime exploration — and the cost of phantom APIs.

Evals & Observability

Three evaluation surfaces. Traces with bodies. The issue lifecycle.

AFK & Autonomous Agents

The Ralph loop, agent fleets, and the HITL→on-rails→AFK gradient.

The Failure Taxonomy

A field guide — memory, reflection, planning, action, system faults.

Worked example — Evals

Pillar III · The Craft

Build the eval surface

Three evaluation surfaces — unit evals on discrete steps, regression suites, and continuous production trace sampling.

Pillar II · The Operating Model

Standardise the eval pipeline

The platform team provides shared eval infrastructure, dashboards, and quality gates that every agent must pass.

Pillar I · The Groundwork

Set the quality bar

Leadership defines acceptable accuracy, latency, and cost thresholds per risk tier, and ties them to the governance framework.