Concepts
The vocabulary, defined.
Plain definitions of the ideas this site runs on. Each links to where the idea is developed in depth.
Agentic Engineering
A delivery model in which AI agents work toward goals — planning, executing and checking with real tools — while humans set intent and govern acceptance.
AI-Native Organisation
The maturity stage where AI ways of working are the default operating model — strategy, data, governance and platforms re-engineered around them, not a programme bolted on.
Architecture Decision Record (ADR)
A short document capturing one significant decision — context, decision, consequences — held in source control and superseded, never edited.
Context Engineering
Curating the conventions, decisions, contracts and constraints an agent sees — the largest lever on whether agentic output is acceptable.
Decision Architecture
Treating architecture as the quality and velocity of decisions — classify by reversibility, decide via advice, capture as records that supersede rather than mutate. Extended to AI, it is the organisation’s decision rights, evidence gates and reversibility.
Delivery Architecture
The discipline — owned by no one in most organisations — that translates business strategy into running outcomes across six handoffs.
Evals as Spec
Treating an executable evaluation suite as the real specification — a criterion two competent reviewers would agree on, made decidable before generation.
Governance-to-Value Ratio
A lens for settling whether a control earns its keep: the decision quality it adds versus the delay it imposes. Most controls never settle the wager.
Harness Architecture
The engineering around the model — scopes, context, and deterministic gates — that turns a capable model into acceptable output. The harness is the product.
Human Acceptance Loop
The supervised loop in which a human accepts or rejects agent-produced work at the points that matter — human-on-the-loop, not autopilot.
Intent Translation
Preserving intent as it moves from strategy through product, architecture, engineering and operations — and the place most of it quietly leaks away.
Provenance Engineering
Recording who decided what, when and on what evidence — an unbroken chain of custody for decisions in human-plus-agent systems.
The Acceptance Gap
The distance between what an agent generates and what a team will actually trust and ship. Once generation is cheap, this gap is where the work lives.
The Agent Gateway
The brokered control point where an AI agent’s access to real systems concentrates — scoped identity, action-level authorization, and audit & reversibility — run on the principle of Least Agency.
The Context Supply Chain
The end-to-end flow by which context reaches an agent, governed like a supply chain on three trust questions — Provenance (where from?), Freshness (still current?), Integrity (uncorrupted?) — because your agent is only as good as its worst context supplier.
The Conviction Gap
The widening distance between what an organisation can now cheaply build and what it has evidence to justify scaling — the judgement problem AI abundance creates.
The Designed Apprenticeship
Deliberately engineering how junior engineers acquire senior judgement — aimed at the evaluative function and taught by making reasoning visible — now that AI has removed the foundational work that used to teach it for free.
The Governing Narrative
What a board actually governs about AI — Portfolio, Reversibility and Competence — run on the principle Evidence Over Theatre; decision architecture at the highest tier of the organisation.
The Kill Rate
The share of AI use cases an organisation deliberately stops — a leading indicator of a healthy portfolio; paired with Default Off, the philosophy that production must be earned, not assumed.
The Living Index
The lifecycle that keeps a retrieval system tracking reality instead of rotting — Freshness, Permissions and Evaluation — producing the outcome everyone wants, Durable Retrieval. Retrieval is a living system, not a launch.
The Margin Floor
The structural gross-margin level an AI product settles toward once inference is a real cost of goods — held down because efficiency gains get reinvested into more usage — and raisable only by design.
The Production Gap
The chasm between an AI pilot that works and an operating model that delivers value — the organisational second gap where most pilots stall before ROI.
Vibe Coding
Generating code by prompting on instinct and accepting what looks plausible — fast, undisciplined, and the thing agentic engineering moves teams beyond.