-
Notifications
You must be signed in to change notification settings - Fork 4
Glossary
The vocabulary, in one place, so no other page has to redefine it.
Terms with a page get a link. Everything else is defined here and nowhere else, which is the point: a term defined in four places drifts in three of them.
Where this glossary and a page disagree, the page wins.
Agentic CoE · The mature form of the AI function. Not a governance board: it works as a factory that builds assets, a governor that enforces guardrails, and an enabler that provides the paved road. See The agent factory.
Agent drift · Performance decaying over time, not because the code changed but because the world did: schemas shifted, behaviour moved, an upstream interface changed. See The three drifts.
Alignment trap · The paralysis of trying to agree a perfect three-year strategy before building one working prototype.
Antifragility · The property of getting stronger after a failure rather than merely surviving it. Operationally: every incident ends in a new regression test or guardrail.
Automation theatre · Activity that looks like progress and delivers nothing. Counting bots launched rather than hours redirected. Demanding an AI headline for an investor call with no operational substance underneath it.
Bot sprawl · The entropy of scaling, where business units copy pilot code into dozens of disconnected unmanaged agents. The opposite of the paved road, and the reason for the registry.
Circuit breaker · A control that automatically degrades an agent, to read-only or draft mode, when a threshold is breached. Distinct from a kill switch, which stops it and requires a person. See Governance as code.
Commando unit, or proto-CoE · The first small cross-functional team: sponsor, architect, product lead, risk lead. Their job is to survive the pilot stage and prove value fast. Their exemption from normal process is deliberate and temporary. See The talent stack.
Composability · Breaking a monolithic agent into specialised components to limit the blast radius of any one error.
Context engineering · Pruning and shaping what an agent is given, to reduce noise and fabrication. The working rule is to aggressively minimise context rather than maximise it.
Customer zero · Using internal teams as the first and harshest users of a new agent, to earn trust before it reaches anyone outside.
Deployment gauntlet · The transition from working pilot to production system, where latency, cost, compliance and drift kill projects. See The deployment ladder.
Determinism axis · A way of reading a use case. High-determinism work has clear rules and fixed outcomes and can run on rails. Low-determinism work turns on nuance, negotiation or judgement and needs supervision built in. See Choosing what to build.
Dignity clause · The requirement that the reskilling path for affected roles is defined before kickoff, not after go-live, and that value reports say where the freed hours went. See Value and the CFO.
Doer to supervisor · The role shift agents cause: from doing the task to supervising the thing that does it. A different skill, not a promotion. See People, power and politics.
Enterprise brain · A semantic layer plus a knowledge graph, so agents read core business concepts consistently across systems rather than re-deriving them.
Enterprise readiness gap · The distance between what a model can do and what your architecture will let it do. Data silos, brittle interfaces, undocumented systems.
FinOps, agentic · Managing unit economics: cost per transaction, caching, and making sure expensive models are reserved for reasoning that is worth it.
Front door, or intake charter · The standard process for reviewing new agent ideas, so the loudest request is not the next project. See Finding the work.
Glue work · The invisible manual labour holding a process together: copying data, chasing follow-ups, reconciling by hand. Usually the first thing an agent absorbs, and the reason adoption can feel like erasure.
Golden dataset · The test set that defines what correct means, built by the domain expert rather than the developer. See The three proof gates.
Guardrail stack · Three layers of defence: policy (rules and never events), technical (filters, validators, enforcement in code), organisational (named owners, accountability, review). See Governance as code.
Hallucination, or confident fiction · A plausible and factually wrong output. In enterprises it usually takes the form of an invented policy, discount, or citation.
Hub and spoke · A central hub builds the platform and standards; business-unit spokes build domain agents on it. See The agent factory.
Maturity cliff · The point where early pilot excitement fades and operational complexity makes risk and cost spike at the same time.
Meaning gap · The gap between a correct output and a trusted outcome. See The framework spine.
Monolithic agent · An anti-pattern: one agent trying to handle retrieval, logic, tone and action at once. See Agent patterns.
Morning routine · The daily ritual for production agents: synthetic tests, then drift checks, then human review, run before the working day. See The three drifts.
Never events · Outcomes the policy layer says an agent must never execute, written specifically enough to test. Not "be careful with pricing" but "never promise a price that is not in the database".
OODA loop · An agent's operational cycle: observe the input, orient on context, decide a plan, act through a tool.
Optimisation trap · Using AI to do the wrong thing slightly faster, rather than removing the need to do it.
Outer loop · Where the agent meets reality: deployment, monitoring, feedback, drift. Distinct from the inner loop of prompting.
Paved road · The standard platform the CoE provides: infrastructure, logging, guardrails, registry. On it you get faster approval and support. Off it you own your own risk, in writing.
Pilot purgatory · Dozens of successful demos and zero agents in production, because nothing crossed the gap of trust, security or value.
Pre-flight checklist · The mandatory gates an agent passes before production. Ten of them, published.
Production gate question · "If this agent gives the right answer to the wrong question, how would you know?" See The framework spine.
Risk tier · One to four, by blast radius rather than by model: read-only, internal write, external or regulated, irreversible or financial. See The framework spine.
Second wave · The phase after pilot success, scaling from a handful of agents to hundreds. Where the maturity cliff appears.
Semantic layer · The component that normalises what data means, so agents do not reverse-engineer business logic out of raw tables.
Seniority ladder · Five levels of agent autonomy and the governance each requires. See The seniority ladder.
Supervisor agent · An agent whose job is judging another agent's output against safety or policy before it reaches anyone. It cannot be weaker than the agent it judges. See Agent patterns.
Synthetic tests · Automated golden scenarios run on a schedule to verify an agent still behaves. The first step of the morning routine.
System of record, audit · The immutable log of prompts, retrieval, reasoning and outcomes. The thing that makes every other control auditable, and the thing most often not built.
Systems of agency · Systems that observe, reason and act. Distinct from systems of record, where humans enter data for storage.
Task-centric API · A wrapper built for an agent, encapsulating validation and error handling, rather than raw table operations. The answer to the API gap. See The agent registry.
Three proofs · Proof of Technology (does it work), Proof of Value (does it matter), Proof of Competence (can we run it, and fix it when it breaks).
Competence, not credibility. The distinction is substantive: credibility asks whether experts trust it, competence asks whether this organisation can operate and remediate it. See The framework spine.
Trust ladder · Four rungs of earned autonomy for one specific agent. See NORTH and the trust ladder.
Two-way door · A reversible decision, which most pilots are. Distinct from a one-way door, which is what a production deployment usually is.
Under-context · A bad decision caused by the agent lacking data it needed, rather than reasoning badly with the data it had.
Under-governance · An agent operating without sufficient constraint, leading to over-reach.
Wild west · Stage one of adoption: high enthusiasm, low governance, shadow AI, no standard patterns. See Why programs stall.
From the practice canon. Developed through client delivery and refined in collaboration; the frameworks stated here are the author's own.
The thinking
Operating model
Frameworks
Governance
Playbooks
Value and people
Reference
In the repository
The courses
Reviewed 2026-08.