Definitions for the Agent Era
Precise definitions for AI Agent Lifecycle, Agentic Delivery, Confirmation Boundary, Evidence Chain, Multi-Agent Systems, and related terms. Stable anchors for citation. Defined by Jearon Wong.
Definitions is a reference index for stable citation anchors. Canonical concept explanations live under /concepts/, and the full entity mesh lives under /concepts/map/. GAIC-derived terms should route from these anchors to white-paper-backed concept or governance pages instead of treating this index as the canonical explanation surface.
Core thesis
AI Agent Lifecycle
AI Agent Lifecycle defines the accountable lifecycle of agent work from intent to accepted outcome.
Not a deployment lifecycle, runtime state lifecycle, manifest lifecycle, or workflow lifecycle. Those systems may be useful. They do not define the accountable lifecycle of agent work.
Agentic Delivery
Agentic Delivery names the missing layer between agent execution and accountable outcomes.
Not prompt engineering, context engineering, or harness engineering. Delivery means carrying human intent through context, planning, confirmation, execution, evidence, review, and accepted outcome.
Execution is not Delivery
Execution proves that an agent can act, while delivery proves that the work reached an accepted outcome with responsibility, evidence, and authority still attached.
Execution is a necessary condition. It is not a sufficient condition for accountable delivery.
Accepted Outcome
An Accepted Outcome is a result that has been reviewed against original intent, constraints, and evidence, then formally accepted.
Not task completion. Not model output. Not a passing evaluation score. Acceptance requires a review record, not merely a visible result.
Accountable Delivery
Accountable Delivery is agent work that stays tied to intent, authority, responsibility, evidence, and review from start to accepted outcome.
Not observable work. Not evaluatable work. Accountable delivery is work that can be traced, reviewed, disputed, and accepted.
Governance primitives
Confirmation Boundary
A Confirmation Boundary is the lifecycle point where autonomous execution becomes authorized responsibility.
Not a UI approval button. A Confirmation Boundary defines what plan is being approved, which scope is covered, who is authorizing, and how that authorization links to the original intent and active constraints.
Evidence Chain
An Evidence Chain is structured proof that agent work can be reviewed, replayed, disputed, and accepted.
Not raw logs. Not traces. An Evidence Chain is structured support for a delivery claim: artifacts that can reconstruct why the work began, what plan was approved, what happened, and why the outcome should be accepted.
Audit Evidence Chain
An Audit Evidence Chain is a responsibility-linked evidence chain for agentic work that connects authority, role, tool action, evidence, outcome, exception, privacy treatment, and remediation closure.
Not raw logs, traces, screenshots, or observability alone. The Agentic AI Auditability & Assurance White Paper 2026 treats audit evidence chains as lifecycle evidence architecture, not as an audit standard or assurance opinion.
Agentic AI Auditability
Agentic AI Auditability is the ability to reconstruct, test, and evidence agentic lifecycle work across authority, responsibility, tools, outcomes, exceptions, privacy treatment, and closure.
Not certification, an audit standard, legal compliance proof, assurance opinion, or vendor ranking. It is the auditability and assurance white paper's public research edition framing.
Agentic Audit Object
An Agentic Audit Object is a proposed review object for agentic work that makes delegated lifecycle activity inspectable through authority, role, tool, evidence, outcome, exception, privacy, and closure fields.
Not a legal liability object, certification criterion, mandatory implementation schema, or audit-procedure template.
Agentic Auditability Readiness Model (AARM)
AARM is the Agentic Auditability Readiness Model from the Agentic AI Auditability & Assurance White Paper 2026, describing L0-L5 readiness states for lifecycle evidence, audit evidence chains, and assurance-planning discussion.
Not a score, benchmark, certification, assurance result, legal compliance proof, procurement tool, or vendor comparison.
Agentic AI Insurability
Agentic AI Insurability is the ability to describe agentic work through lifecycle evidence, responsibility mapping, bounded risk objects, and claim-reviewable records for risk-transfer discussion.
Not insurance advice, a coverage opinion, insurer acceptance, coverage-ready status, underwriting-ready status, certification, or a guarantee that any system is insurable.
Agentic Insurability Objects (AIO)
Agentic Insurability Objects are analytical objects from the Agentic AI Insurability & Risk Transfer White Paper 2026 that separate insured legal subject, agentic risk object, authority, responsibility, evidence, loss reconstruction, dependency, aggregation, and dispute-readiness questions.
Not insurer product requirements, policy terms, legal liability objects, certification criteria, or a mandatory implementation schema.
Agentic Insurability Reasoning Model (AIRM)
AIRM is the Agentic Insurability Reasoning Model from the Agentic AI Insurability & Risk Transfer White Paper 2026, a non-scoring vocabulary for evidence visibility, claims review, underwriting discussion, and dispute readiness.
Not an actuarial score, insurer acceptance, coverage guarantee, underwriting standard, claims approval guide, certification, vendor score, or procurement benchmark.
Insured Legal Subject
An Insured Legal Subject is the person or organization whose risk-transfer relationship must remain separate from the agentic system, work unit, tool, model, or workflow being analyzed.
Not a liability determination, coverage opinion, insured-status opinion, or conclusion that a policy applies.
Agentic Risk Object
An Agentic Risk Object is the bounded agentic work unit, action path, dependency, evidence chain, or loss-relevant lifecycle object being evaluated for risk-transfer analysis.
Not the insured party, not a legal subject, not a standalone coverage trigger, and not a claim that a system is insurable.
Claim Evidence Chain
A Claim Evidence Chain is the lifecycle evidence needed to reconstruct authority, action, loss event, dependency, remediation, dispute posture, and boundary risk for claim review.
Not claims approval guidance, a payment guarantee, legal causation proof, settlement advice, or an insurer-required form.
Lifecycle-Responsibility-Linked Agent Work
Lifecycle-responsibility-linked agent work is agent work whose intent, authority, responsibility, tool actions, evidence, review, accepted outcome, exception handling, and closure remain connected.
Not task completion, autonomous execution, generic observability, or a claim that a system is audit-ready by default.
Semantic Loss
Semantic Loss is the degradation of intent, constraints, responsibility, and evidence across lifecycle handoffs.
Not hallucination. Not context window overflow. Semantic Loss is the failure mode where meaning, authority, and constraints are silently dropped during agent work, often without any single step appearing wrong.
Multi-Agent Lifecycle Governance
Multi-Agent Lifecycle Governance is the multi-agent form of Agentic Delivery: governing responsibility, authorization, evidence, and outcome acceptance across agents, roles, projects, and lifecycle stages.
Not multi-agent coordination. Not orchestration. Multi-Agent Lifecycle Governance requires authority, evidence, and accepted outcome, not only task routing.
Intent Drift
Intent Drift is the gradual separation between the original human objective and the direction an agent system actually follows.
Not model hallucination. Not factual error. Intent Drift is a lifecycle failure where the system lacks a stable way to preserve, update, and verify intent over time.
Context Drift
Context Drift is the loss of fit between the context an agent uses and the actual state of the work.
Not a context window limitation. Context Drift happens when summaries compress reasoning, old rules remain active after requirements change, or constraints are present but not weighted correctly.
Lifecycle Responsibility Consensus
Lifecycle Responsibility Consensus is the orchestration-layer mechanism that aligns human intent, role authority, agent execution, evidence, review, and accepted outcome into a traceable delivery relationship.
Lifecycle Responsibility Consensus is not ordinary agent routing, workflow continuation, or human approval. It describes how the orchestration layer aligns responsibility, execution, evidence, review, and acceptance into one traceable delivery relationship.
System architecture
MPLP
MPLP is the lifecycle protocol path for making Agentic Delivery explicit, governable, and auditable.
MPLP does not equal Agentic Delivery. MPLP is the protocol path inside the Agentic Delivery category, not the category itself.
Protocol Engineering
Protocol Engineering is the discipline of making the critical states and transitions of agent work explicit enough to be implemented, checked, and shared.
Not application logic. Not prompt rules. Not observability. Protocol Engineering defines what must be true before agents act and what must remain traceable after they act.
Lifecycle Role Decomposition
Lifecycle Role Decomposition translates human work roles into lifecycle responsibility boundaries that agent systems can execute, confirm, trace, roll back, and accept.
Not renaming PM, Architect, Developer, Reviewer, or QA as agents. Lifecycle Role Decomposition decomposes the responsibility behind those roles into lifecycle objects so agent systems can operate them with accountability.
Lifecycle-Governed Agent Workflow
A Lifecycle-Governed Agent Workflow is a workflow model in which a human-readable work process is interpreted through lifecycle protocol and generated as a governed agent workflow.
Not a node graph. A Lifecycle-Governed Agent Workflow carries role boundaries, confirm gates, trace obligations, rollback points, and delivery states, not only execution edges.
Market corrections
Multi-Agent Systems
A multi-agent system is an agent architecture in which responsibility for work is separated across distinct lifecycle roles, not merely an architecture with more than one agent.
Multi-agent is not multi-agent count. Two agents sharing a context window without responsibility separation is not a multi-agent system. It is a parallel execution pattern.
Agent Orchestration
Agent Orchestration is the coordination of multiple agents for execution, directing which agent runs next, under what conditions, and with what inputs.
Agent Orchestration is not lifecycle governance. Orchestration coordinates execution. Governance defines authority, accountability, evidence, and accepted outcome: conditions that persist beyond any single execution run.
Human-in-the-Loop (HITL)
Human-in-the-Loop (HITL) is a design pattern in which a human is positioned to observe or approve actions at one or more points in an agent workflow.
HITL is not governance. A human can be present and still lack the information needed to authorize responsibly. Lifecycle governance requires that the human's confirmation carries explicit scope, plan context, evidence obligation, and return condition.
AI Agent Governance
AI Agent Governance is the set of authority, boundary, confirmation, evidence, and review conditions that make delegated agent work accountable.
Not permission management, access control, or monitoring. Those are necessary conditions. Governance also requires lifecycle continuity: authority must attach to intent, plans must carry constraints, confirmations must be recorded, and evidence must survive after execution.
Delivery Standard
The Delivery Standard is the set of conditions under which AI agent work counts as accountable delivery: scope, authority, evidence, review, and accepted outcome.
Not an output standard. Not an evaluation score. The Delivery Standard asks whether the work can be traced from intent to accepted outcome with responsibility still attached.