Glossary

Short definitions of the terms used across the lab. Each links to the step where the idea is put to work.

Search the glossary

ABACstep 5: Attribute-based access control: permissions decided from attributes of the user, the resource and the situation, such as department, classification or time.
Act with approvalstep 4: The third autonomy level: the agent executes an action only after a human confirms it.
Adequacy decisionstep 5GDPR Art. 45(1)(opens in a new tab): A Commission decision that a country outside the EU protects personal data adequately, so data may go there without further authorisation.
Agentstep 1: A system that can plan and act across tools, under constraints and policies. Unlike a copilot, it carries out the work itself.
AgentOpsstep 8: The operational discipline for agents: evaluation, testing, release gates and monitoring, closer to SRE than to prototyping.
AI Actstep 4AI Act(opens in a new tab): The EU's Artificial Intelligence Act, Regulation (EU) 2024/1689: rules for AI by risk, from banned practices to high-risk duties and transparency.
AI FinOpsstep 11: Cost governance for AI: budgets by use case, quotas, model routing, caching and cost monitoring tied to business outcomes.
AI literacystep 10AI Act Art. 4(opens in a new tab): The skills and understanding to use AI knowingly. Providers and deployers take measures to support it among the people who run and use their AI.
All other bucketstep 6: The last line of an app-computed breakdown: the largest categories are listed, and everything else is counted in one bucket, so the parts add up to the total.
Allow-liststep 9: An explicit list of what is permitted, such as the tools an agent's identity holds connectors for, or the destinations data may go to. Anything not on it is refused.
Approval gatestep 6: A point where the agent must wait for a human before an action runs; used for irreversible or high-impact actions.
Audit logstep 4: An immutable record of each step, tool call, decision and approval: who requested it, who approved it, with which parameters, and what happened.
Autonomy levelstep 4: How far an agent may go on its own: read-only, recommend, act with approval, or act autonomously.
Blast radiusstep 2: How much damage a wrong action can do. Good first use cases have a small, well-bounded blast radius.
Canary deploymentstep 8: Releasing a new agent version to a small group first, such as one service team, with automatic rollback if quality drops.
Context packetstep 1: The first output of the incident agent: blast radius, recent changes, top error signatures and recommended next checks.
Copilotstep 1: An assistant inside an application; the human remains the executor.
Cyber Resilience Actstep 5CRA Art. 71(2)(opens in a new tab): The EU regulation on cybersecurity for products with digital elements made available on the market; it applies in full from 11 December 2027.
Data Actstep 5Data Act Art. 23(opens in a new tab): The EU regulation on fair access to and use of data. For agents it matters mostly for the right to switch cloud providers.
Data minimizationstep 7: Giving the agent only the data its task needs, with personal data masked where possible.
Deny by defaultstep 4: A policy stance: anything not explicitly allowed is forbidden.
Deployerstep 3AI Act Art. 3(4)(opens in a new tab): Under the AI Act, whoever uses an AI system under its own authority, other than for personal, non-professional use.
Describe, never computestep 6: A rule for agent output: the application computes every figure and names every system, and the model only describes them. A checker drops any sentence that goes further.
Digital Omnibus on AIstep 12Digital Omnibus on AI(opens in a new tab): Regulation (EU) 2026/1744, in force since 27 July 2026, which simplified the AI Act and moved its high-risk dates to December 2027 and August 2028.
DORAstep 9DORA Art. 2(1)(opens in a new tab): The EU regulation on digital operational resilience for the financial sector: ICT risk, incident reporting and third-party risk for banks, insurers and other financial entities.
DPIAstep 4GDPR Art. 35(1)(opens in a new tab): Data protection impact assessment: an assessment, made before the processing starts, of processing likely to result in a high risk to people.
Egress controlstep 9: Limiting where data may leave the company for, so an agent cannot send it to an outside address.
Fallback modestep 6: Degrading gracefully to recommending instead of acting when the agent is unsure or the intent is unclear.
FRIAstep 4AI Act Art. 27(1)(opens in a new tab): Fundamental rights impact assessment: required before first use of certain high-risk AI systems by public bodies, providers of public services, and credit and insurance deployers.
GDPRstep 4GDPR Art. 2(1)(opens in a new tab): The General Data Protection Regulation, Regulation (EU) 2016/679: the EU's rules for processing personal data.
General-purpose AI modelstep 5AI Act Art. 3(63)(opens in a new tab): An AI model that can competently perform a wide range of distinct tasks and be built into many systems, such as the large language model behind an agent.
Golden tasksstep 8: Known scenarios, often replayed past incidents, that every release of an agent must pass: a regression suite.
Governance as codestep 12: Policies expressed as machine-enforced rules, checked before anything runs, instead of documents.
High-risk AI systemstep 4AI Act Art. 6(opens in a new tab): Under the AI Act, AI used in the areas of Annex III or as a safety component of certain products. It carries duties such as risk management, logging and human oversight.
Human oversightstep 6AI Act Art. 14(opens in a new tab): The ability of people to understand, monitor, override and stop a high-risk AI system while it is in use.
Idempotencystep 8: The property that doing something twice has the same effect as doing it once, which makes retries safe.
Indirect prompt injectionstep 9: Instructions hidden in content the agent retrieves, such as a log line, ticket or web page, written to make it act against its purpose.
Lawful basisstep 4GDPR Art. 6(1)(opens in a new tab): One of the six grounds on which the GDPR allows personal data to be processed, such as a contract, a legal obligation or legitimate interests.
Least privilegestep 9: Giving an identity only the permissions its current task needs, scoped to the right service and environment.
Major ICT-related incidentstep 9DORA Art. 19(4)(opens in a new tab): Under DORA, an ICT incident at a financial entity serious enough that it must be reported to the authority, starting with an initial notification within hours.
MTTRstep 11: Mean time to resolve: from an incident starting to the service being healthy again.
NIS2step 9NIS2 Art. 2(1)(opens in a new tab): The EU directive on a high common level of cybersecurity: risk measures and incident reporting for essential and important entities in its sectors.
Personal data breachstep 9GDPR Art. 4(12)(opens in a new tab): A breach of security leading to the accidental or unlawful destruction, loss, alteration or unauthorised disclosure of, or access to, personal data.
Policy enginestep 6: The component that checks every proposed tool call against the rules before it executes, and explains its decision.
Processorstep 5GDPR Art. 28(3)(opens in a new tab): Someone who processes personal data on your behalf, such as a model provider receiving your logs. It needs a contract with the content the GDPR lists.
Prototype zoostep 5: The article's name for agents each built differently, in different repositories with different logging and permissions.
Provenancestep 7: Where an answer came from: the documents and versions it was based on, cited so they can be checked.
Provider (AI Act)step 3AI Act Art. 3(3)(opens in a new tab): Whoever develops an AI system, or has one developed, and puts it on the market or into service under its own name.
RACIstep 3: Responsible, Accountable, Consulted, Informed: a matrix of who does what for each decision.
RAGstep 5: Retrieval-augmented generation: the model answers from documents retrieved for the question, rather than from memory alone.
RBACstep 5: Role-based access control: permissions attached to roles, and roles to people or services.
Red-teamingstep 9: Deliberately attacking your own agent with injection, exfiltration and jailbreak attempts to find weaknesses first.
Register of informationstep 5DORA Art. 28(3)(opens in a new tab): The register a financial entity keeps under DORA of all its contracts for ICT services from third-party providers.
Runbookstep 7: A written procedure for handling a known kind of incident; the agent follows it and cites it.
Service accountstep 5: A non-human identity for an agent, with its own constrained permissions.
Significant incidentstep 9NIS2 Art. 23(3)(opens in a new tab): Under NIS2, an incident that has caused or can cause severe operational disruption or financial loss, or considerable damage to others. It must be reported.
Standard contractual clausesstep 5GDPR Art. 46(2)(c)(opens in a new tab): Clauses adopted by the Commission that give a legal basis for transferring personal data outside the EU.
Step budgetstep 6: A cap on how many tool calls an agent may make in one session, so a confused agent cannot run away.
Typed tool schemastep 6: The structured definition of a tool's inputs; calls missing a required parameter are refused before they run.
Untrusted contentstep 7: Anything the agent retrieves that someone outside the team could have written. It is read, never obeyed.