Glossary
Agentic AI, in plain terms
The vocabulary of agentic AI, defined dryly: harness, agents, tool use, retrieval, and the sovereignty terms that decide who ends up owning the system. No hype, no em-dashes.
Agentic core
- Agent
- A system that takes a goal, plans steps toward it, acts through tools, observes the results, and repeats until done. An agent is a model plus a harness plus tools; none of the three alone is an agent.
Related: Agentic AI · Harness · Tool Use · Orchestration
- Agentic AI
- AI that pursues a goal over multiple steps, choosing its own actions and tools along the way, rather than answering a single prompt. The shift from generate text to get something done. Autonomy is a spectrum, not a switch; most useful agentic systems keep a human at the decisive points.
Related: Agent · Harness · Tool Use · Human-in-the-Loop
- Agentic Engineering
- The discipline of building systems where a model acts, not just answers. It moves the engineering effort off the prompt and onto the three things that decide outcomes: the tools you expose, the context you assemble, and the condition on which the loop stops. Prompting is a skill; agentic engineering is an architecture practice.
Related: Agentic Harness · Loop Engineering · Context Engineering · Eval · Agentic AI
- Agentic Harness
- The production-grade harness: not just the loop, but everything that makes an agent safe to run unattended. Tool permissions, sandboxing, retries, budget and step limits, logging, and the termination conditions. The harness is where an agent stops being a demo and becomes a system you can operate.
Related: Harness · Guardrails · Loop Engineering · Agentic Engineering
- Context Engineering
- Deciding what information reaches the model, in what form, at what moment. The successor discipline to prompt engineering: less about wording, more about assembling the right facts, tools, and history into a limited context window. Often the difference between a flaky agent and a dependable one.
Related: Context Window · Retrieval-Augmented Generation (RAG) · Agentic Engineering · Harness
- Context Window
- The maximum amount of text a model can consider at once, prompt and response together, measured in tokens. Everything the model knows in the moment must fit inside it. Run out of room and the earliest content falls away, so what you put in and leave out is a design decision.
Related: Context Engineering · Embeddings · Retrieval-Augmented Generation (RAG)
- Eval
- A repeatable test that measures whether a model or agent does its job, on cases you care about, with a score you can track over time. Without evals you are not engineering an AI system, you are guessing about one.
Related: Guardrails · Provenance
- Guardrails
- Constraints that keep a model's behaviour inside acceptable bounds: input filters, output validation, allowed-tool lists, policy checks. Guardrails do not make a model correct; they make its failures visible and contained.
Related: Eval · Human-in-the-Loop · Prompt Injection
- Harness
- The code that surrounds a language model and turns it into a working system: the loop that calls the model, runs its chosen tools, feeds results back, and decides when to stop. The model reasons; the harness does everything else. Most of an agent's reliability lives here, not in the prompt.
Related: Agentic Harness · Agentic AI · Agent · Loop Engineering · Tool Use
- Human-in-the-Loop
- A design where a person approves, corrects, or overrides the system at defined points instead of letting it run unattended. In regulated and high-stakes work this is not a fallback; it is the control that keeps judgment inside the organisation.
Related: Guardrails · Cognitive Debt · Agentic AI
- Loop Engineering
- Designing the agent's control flow deliberately: how many steps it may take, what it does on a failed tool call, when it escalates to a human, and above all when it stops. Termination is a correctness property, not a detail. A loop with no robust exit is how you get runaway agents and token bills that scale with confusion.
Related: Harness · Agentic Harness · Agentic Engineering · Guardrails · Human-in-the-Loop
- Multi-Agent System
- Several agents working on one problem, each with a role, passing work between them. Powerful when a task splits cleanly into independent parts, wasteful when it does not. More agents means more coordination cost, so the split has to earn it.
Related: Agent · Orchestration
- Orchestration
- Coordinating multiple models, tools, or agents into one reliable workflow: what runs when, what happens on failure, how state passes between steps. As soon as a system has more than one moving part, orchestration is where correctness is won or lost.
Related: Harness · Multi-Agent System · Model Router
- Tool Use
- The mechanism by which a model invokes external functions, APIs, or code instead of only producing text. The model emits a structured call; the harness runs it and returns the result. Tool use is what lets an agent touch the real world: query a database, send an email, open a file.
Related: Harness · Model Context Protocol (MCP) · Agent · Guardrails
Protocols & retrieval
- AI Gateway
- A single controlled entry point between your applications and model providers, where routing, logging, rate limits, cost tracking, and provider swaps live. The gateway is where sovereignty over models becomes an architectural fact instead of a slide.
Related: Model Router · Sovereign AI · Provenance · Vendor Lock-in
- Embeddings
- Numeric vectors that place text, images, or code in a space where nearness means similar meaning. They power search, clustering, and retrieval. They are also the quietest lock-in: change the embedding model and every stored vector has to be recomputed.
Related: Retrieval-Augmented Generation (RAG) · Vendor Lock-in · Knowledge Graph
- GraphRAG
- Retrieval that walks a knowledge graph instead of ranking loose text chunks, so the model gets connected facts and their relationships, not just nearby sentences. Stronger when the answer depends on how entities relate, weaker when a flat lookup would do.
Related: Retrieval-Augmented Generation (RAG) · Knowledge Graph · Embeddings
- Knowledge Graph
- A model of a domain as entities and the typed relationships between them, queryable as a graph. It gives an AI system a structured memory that can be inspected and corrected, rather than facts smeared across weights.
Related: GraphRAG · Retrieval-Augmented Generation (RAG) · Provenance
- Model Context Protocol (MCP)
- An open standard for connecting models to tools and data through a uniform interface, so a capability built once works across any compatible client. MCP is to agent tooling what a driver model was to hardware: it turns bespoke integrations into a plug.
Related: Tool Use · AI Gateway · Harness
Read: Compliance by Design (DE) → · Your Coding Agent Is Just a While-Loop →
- Model Router
- Logic that sends each request to the most suitable model by cost, capability, latency, or jurisdiction, behind one interface. A router turns which model from a lock-in decision into a runtime choice you can change without touching callers.
Related: AI Gateway · Sovereign AI · Orchestration
- Retrieval-Augmented Generation (RAG)
- Fetching relevant documents at query time and placing them in the context window so the model answers from your data, not only its training. The standard way to ground a model in current, private, or citable facts without retraining it.
Related: Embeddings · GraphRAG · Context Engineering · Knowledge Graph
Sovereignty & governance
- Adoption Rings
- The Adopt, Trial, Assess, Hold classification popularised by the technology radar. A way to say not just what exists but what you should do about it. Useful only if the placement is earned by evidence rather than assigned by opinion.
Related: Eval · Model Router
Read: AI Radar →
- AI-Native Transformation
- Rebuilding how an organisation works around AI, rather than bolting AI onto how it already works. The tell is where the judgment sits: a bolted-on rollout buys tools and hopes for output; a native transformation redesigns the decision paths, the review, and the architecture that keeps those decisions explainable. Done badly, it is the fastest way to accumulate Cognitive Debt.
Related: Cognitive Debt · Sovereign AI · Cognitive Debt Maturity Model (CDMM) · Agentic Engineering
Read: Cognitive Debt: the framework → · Das Cognitive Debt Maturity Model (DE) →
- Cognitive Debt
- The slow erosion of an organisation's judgment when AI systems produce decisions no one can any longer explain or check. Like technical debt it is invisible until the interest comes due, as liability, compliance gaps, and lost competence.
Related: Sovereign AI · Cognitive Debt Maturity Model (CDMM) · AI-Native Transformation · Provenance
Read: Cognitive Debt: the framework → · Das Cognitive Debt Maturity Model (DE) → · Idea Assessor →
- Cognitive Debt Maturity Model (CDMM)
- A proprietary framework that makes an organisation's AI judgment-readiness measurable, from Level 0 (black-box dependent) to Level 4 (sovereign). Each level is climbed by one architecture decision, not by adding models. A way to see the debt before the interest arrives.
Related: Cognitive Debt · Sovereign AI
Read: Cognitive Debt: the framework → · Das Cognitive Debt Maturity Model (DE) →
- EU AI Act
- The European Union's risk-based regulation of AI systems, with obligations that rise with the risk class. For builders the practical core is Article 12: high-risk systems must keep logs that make their operation traceable. Provenance stops being optional.
Related: Provenance · Sovereign AI · Cognitive Debt
- Prompt Injection
- An attack where hostile text in a model's input hijacks its instructions: hidden commands in a web page, a document, a tool result. The agentic version is worse, because an agent can act on the injected instruction. Guardrails and least-privilege tools are the defence.
Related: Guardrails · Tool Use · Agentic AI
- Provenance
- The recorded origin of an AI output: which model, which prompt version, which inputs, which raw response. Provenance is what lets you reconstruct a decision months later. An assessment you cannot reconstruct is a gut feeling with a better interface.
Related: Cognitive Debt · Sovereign AI · Eval · EU AI Act
- Sovereign AI
- AI architecture the organisation actually owns: auditable, exit-capable, resilient, EU-compatible, without vendor lock-in. Not on-premise for its own sake, but control over model, data, logic, and liability. The structural answer to Cognitive Debt.
Related: Cognitive Debt · AI Gateway · Vendor Lock-in · Provenance
Read: Cognitive Debt: the framework → · Compliance by Design (DE) →
- Vendor Lock-in
- The state where switching AI provider is so costly you effectively cannot, whether through proprietary APIs, embeddings, fine-tunes, or data gravity. Lock-in is not always wrong, but unpriced lock-in is a bill your successor pays.
Related: Sovereign AI · Model Router · Embeddings