Glossary

Key terms, clearly explained

The most important terms around agentic AI and enterprise architecture — deliberately concise. The depth is in the seven missions.

AgentMission 004 →
An AI system that does not merely respond but plans autonomously, selects tools and carries out actions. It chooses part of its path on its own — unlike rigidly programmed software.
Copilot
Assistance that offers suggestions to a person without replacing them. The decision and the execution remain with the human.
Chatbot
A dialogue system that predominantly responds. As a rule, it does not act autonomously within business systems.
Control PlaneMission 001 →
The cross-cutting control layer above all architecture levels: identity, policy, security, quality, observability, model selection, cost and governance. Key point: it is not the individual agent — the Control Plane is the platform.
Context EngineeringMission 002 →
Deliberately providing the right, up-to-date, authorized and traceable context — often more important than fine-tuning prompts.
RAG (Retrieval-Augmented Generation)Mission 002 →
Targeted retrieval of relevant information instead of writing everything into the prompt. Connects a model with curated knowledge.
ProvenanceMission 002 →
The traceable origin of a piece of information or an artifact — the prerequisite for tracing a statement back to its source.
Tool CallMission 003 →
The concrete invocation of a tool by an agent (e.g. an API or a search function). Tool Calls are the points where an agent has real-world effect — and must therefore be logged.
Tool ContractMission 003 →
The precise description of what a tool can do, where its limits lie and what side effects it has — unambiguous even for a non-deterministic system.
MCP (Model Context Protocol)Mission 003 →
An open standard for connecting AI applications with context sources and tools. It standardizes the integration — but replaces neither authorization nor data quality.
A2A (Agent2Agent)Mission 003 →
A protocol for collaboration between agents. Maturity and version should be checked at every mention.
IdempotencyMission 003 →
The property of a call that produces no duplicate effect even when repeated — important when an agent retries a step.
Non-Human IdentityMission 004 →
A dedicated, verifiable machine identity for an agent — instead of a shared collective account. Makes actions attributable.
Least PrivilegeMission 004 →
The principle of minimal rights: an agent receives only the permissions it genuinely needs for its purpose — separated into reading, proposing and executing.
DelegationMission 004 →
The traceable transfer of human authority to an agent — with clear limits, time-bound and attributable to a responsible person.
Human-in-the-LoopMission 004 →
A mandatory human approval at defined points before an agent triggers a consequential action. Keeps responsibility with the human.
Kill SwitchMission 004 →
A defined way to halt ongoing agent activity immediately. Must be tested, not merely planned for.
EvalMission 005 →
An ongoing assessment of the functional quality of an AI system against curated test cases — classic tests alone are not enough for probabilistic systems.
Golden DatasetMission 005 →
A collection of stable reference cases against which quality is compared and regressions are detected.
ObservabilityMission 005 →
The end-to-end observability of a system — for agents down to the individual tool call, because that is where real harm arises.
Model GatewayMission 006 →
A controlled point for model selection and switching. Keeps model dependencies manageable without abstracting everything away above the architecture.
FinOps for AIMission 006 →
The cost and value management of AI workloads — measured by cost per business outcome, not just per token.
AI-SBOMMission 007 →
A machine-readable bill of materials of the components and dependencies of an AI/ML system — the basis for provenance and a secure supply chain.
Architecture Fitness FunctionMission 007 →
An automatically verifiable architecture rule. Maintains quality as the volume of generated code grows.