The complete agentic design vocabulary. Every term from containment and HITL to failure choreography and trust chains, defined in plain language with design implications.
Glossary
Agentic AI systems are being deployed across healthcare, fintech, and cybersecurity right now. Engineers have the vocabulary to build them. Designers often don’t – and that gap costs organisations dearly when systems fail, governance questions arise, or regulators ask for documentation.
This is the reference you need. Every term used in agentic system design, translated from engineering language into design language. Use Ctrl+F to search. Bookmark this page and come back when you encounter a term in a meeting you don’t recognise.
Agent Architecture
AGENT: A software component that perceives its environment, makes decisions, and takes actions to achieve a goal with some degree of autonomy.
Design implication: every agent needs a defined scope of authority. What can it decide alone? What must it escalate?
AGENTIC SYSTEM: A system composed of one or more agents working together, often with varying levels of autonomy and human oversight.
Design implication: designing an agentic system means covering coordination, containment, failure recovery, and governance – not just the interface.
ORCHESTRATOR: A central agent or process that coordinates the work of other agents, deciding who does what and in what order.
Design implication: the orchestrator’s routing logic is a design deliverable, not an engineering default.
SUB-AGENT: An agent that performs a specific task as part of a larger system, directed by an orchestrator.
Design implication: each sub-agent needs an agent card documenting its inputs, outputs, allowed actions, and risk level.
HANDOFF: The point where one agent passes control, context, or responsibility to another agent or to a human.
Design implication: context must be explicitly preserved at every handoff. What data passes? What gets lost? What does the receiving party need to know?
TOOL USE: When an agent calls external tools such as APIs, databases, or calculators to gather information or take action.
Design implication: every tool expands the agent’s capability and its risk surface. All tool use must be reflected in the containment boundary.
CONTEXT WINDOW: The amount of information an AI model can process at once. Limited context windows mean agents can lose track of earlier parts of a task.
Design implication: design for context loss. What happens when the agent can’t remember what it decided ten steps ago?
HALLUCINATION: When an AI model generates confident but factually incorrect output.
Design implication: critical in safety-sensitive contexts. Design verification steps, confidence scoring, and flag-not-edit patterns to prevent hallucination from propagating through the system.
Coordination Patterns
SEQUENTIAL HANDOFF: Agents work in a chain: Agent A completes, passes to Agent B, then Agent C. Simple but fragile – if one fails, the chain breaks.
Design implication: explicit handoff contracts and failure paths at every link.
PARALLEL COORDINATION: Multiple agents work simultaneously on different aspects of the same task, results are aggregated.
Design implication: design for partial failures. What happens when three agents return results but one fails?
HUB-AND-SPOKE: A central orchestrator routes tasks to specialist agents based on the situation.
Design implication: the orchestrator’s routing logic is a design deliverable. How does it decide who gets what task?
CONDITIONAL ROUTING: The path through the system depends on conditions such as risk score, user input, or external data.
Design implication: map every branch. Every conditional path needs a failure path.
DYNAMIC ARCHITECTURE: Agents can spawn new agents or dissolve at runtime based on need.
Design implication: the most complex pattern. Governance and containment must account for agents that don’t exist yet at design time.
Human Oversight
HUMAN-IN-THE-LOOP (HITL): A human must approve every decision before the system acts. Maximum oversight, minimum speed. Required for irreversible, high-risk decisions.
Design implication: approval gate UI that gives the human reviewer enough context to actually decide – not just a confirm button.
HUMAN-ON-THE-LOOP (HOTL): The system acts, but a human monitors and can intervene.
Design implication: dashboard design, alert thresholds, interrupt controls. The human watches, and can stop the agent.
HUMAN-OUT-OF-THE-LOOP (HOOTL): The system acts fully autonomously. Appropriate only for low-risk, reversible actions.
Design implication: HOOTL is never the default in regulated environments. Justify every instance explicitly.
APPROVAL GATE: A designed pause point where the system waits for human confirmation before proceeding.
Design implication: show the agent’s reasoning, confidence level, and the consequence of each choice. The human needs enough information to genuinely decide.
KILL SWITCH: An immediate shutdown of an agent or system. The last resort when all other containment has failed.
Design implication: accessible, unambiguous, tested. Must work when everything else has failed.
OVERRIDE: A human’s ability to reverse, change, or stop an agent’s decision.
Design implication: override must be technically possible and interface-accessible. Designed in from the start, not bolted on.
Containment and Safety
CONTAINMENT: The designed boundaries around what an agent can and cannot do. Not a safety net, a primary design decision made before a line of code is written.
CONTAINMENT BOUNDARY: The explicit line between autonomous action (green zone), action with notification (amber zone), and action requiring human approval (red zone).
FAILURE MODE: A specific way the system can fail. Good design identifies failure modes before deployment, not in production.
FAILURE CHOREOGRAPHY: The deliberate, designed sequence of events when something goes wrong – from detection through rollback, escalation, user communication, and audit preservation. Not error handling. Intentional orchestration.
GRACEFUL DEGRADATION: When a system loses capability but continues functioning at a reduced level rather than failing completely.
Design implication: what does reduced level look like? What does the user see and what can they still do?
CASCADE FAILURE: When one failure triggers another in a chain reaction. The most dangerous pattern in multi-agent systems.
Design implication: design firebreaks, points where the cascade is designed to stop.
DATA IMMUTABILITY: Records that must be preserved exactly as produced and cannot be edited. The flag-not-edit pattern: users annotate rather than modify the original AI output, preserving the audit trail.
RECOVERY PATH: What the system does after a failure: rollback (undo the action), compensation (make it right), escalation (get help), or graceful degradation (do less safely).
Governance and Design Vocabulary
GOVERNANCE BY DESIGN: Embedding governance decisions at the interface level through design choices like data immutability, confidence scoring, and containment boundaries – rather than through policy documents alone.
TRUST CHAIN: A sequence of decisions where each layer depends on the one before being accurate. If the first link is wrong, everything downstream is wrong.
Design implication: make trust chains visible and traceable.
CONFIDENCE SCORE: A numerical expression of how certain the AI is about its output. Used to calibrate user trust and determine when human review is needed.
ACCOUNTABILITY MAPPING: A documented answer to who is responsible at each decision point in the agent’s operation: the user, the designer, the operator, or the company.
CONSENT DESIGN: Designing the scope of user consent for autonomous agent actions – what the user actually understood they were agreeing to, not just what the terms of service say.
FLAG-NOT-EDIT: A design pattern where users annotate AI output rather than modifying it, preserving the original for audit purposes. Particularly critical in clinical and legal contexts.
INCENTIVE ANALYSIS: Examining what behaviour an AI system rewards and punishes, including unintended second-order consequences at a population level.
SERVICE BLUEPRINT: A diagram showing the entire service across layers: frontstage (what the user sees), backstage (what agents do), and support (what systems enable it). The primary design artefact for agentic systems.
Frequently Asked Questions
Agentic design is the discipline of designing autonomous AI systems – defining what agents are allowed to do, how they coordinate, what happens when they fail, and how human oversight is maintained. It is distinct from UX or interface design because the most consequential decisions happen before any interface is built.
HITL requires human approval before every agent action. HOTL allows the agent to act while a human monitors and can intervene. HOOTL allows fully autonomous operation with no real-time human involvement. The appropriate level depends on the risk, reversibility, and regulatory context of each action.
A containment matrix maps every agent action to one of three zones: green (the agent acts autonomously), amber (the agent acts but notifies a human), or red (the agent cannot act without explicit human approval). It is produced during the design phase, before implementation, to define the scope of agent authority.
Failure choreography is the deliberate, designed sequence of events that occurs when an AI system fails – from detection through to recovery. It covers rollback, escalation, user communication, and audit trail preservation. It is distinct from error handling because it is intentionally designed before deployment, not added reactively.
Article 14 of the EU AI Act requires meaningful human oversight for high-risk AI systems. Designing that oversight – the approval gates, interrupt controls, and monitoring interfaces that enable effective human intervention – is a design responsibility. Designers on high-risk AI systems in regulated industries need to understand and apply these requirements directly to their design decisions.
Join the Agentic Design Community
Want to use this vocabulary in practice?
The Agentic Design Academy teaches every concept on this page through real projects.
