AgentOps Glossary

AgentOpsObservability

AgentOps is the set of operational practices for running autonomous AI agents in production: observability, monitoring, cost control, reliability, security and governance. It extends MLOps and LLMOps to systems that plan, call tools and act over many steps with limited human input.

CheckpointingOrchestration

Checkpointing saves an agent’s state at points during execution so a run can be paused, resumed, rolled back or recovered after failure. It underpins durable long-running agents and human approval steps, letting work continue without restarting from scratch.

EU AI ActGovernance

The EU AI Act is the European Union’s risk-based regulation of artificial intelligence. It phases in obligations by risk tier, including transparency duties under Article 50 that become enforceable on 2 August 2026 and affect how autonomous agents must disclose AI involvement.

Excessive agencySecurity

Excessive agency is the risk that an agent holds more permissions, autonomy or tool access than its task requires. When exploited or triggered by error, broad privileges let a single bad decision cause outsized damage, such as deleting data or moving money.

GuardrailsHuman Oversight

Guardrails are runtime controls that constrain what an agent may receive, produce or do. Input guardrails filter prompts, output guardrails check responses, and behavioral guardrails block disallowed tool actions — together forming a safety layer around the model.

HallucinationEvals

A hallucination is output that is fluent and confident but factually wrong or unsupported. In agents it is especially dangerous because a hallucinated fact can trigger a real tool action, so grounding, verification and evals are used to detect and reduce it.

HOOTLHuman Oversight

Human-out-of-the-loop (HOOTL) describes fully autonomous operation where an agent acts without real-time human involvement. It offers maximum throughput but the highest risk, so it is reserved for low-stakes, well-bounded tasks with strong guardrails and after-the-fact audit.

Human-in-the-loop (HITL)Human Oversight

Human-in-the-loop means a person must approve critical actions before an agent executes them. It maximizes control and accountability at the cost of throughput, and is often required for high-risk steps under emerging AI governance rules.

Human-on-the-loop (HOTL)Human Oversight

Human-on-the-loop means a person supervises an agent and can intervene or stop it, but does not approve each action in advance. It trades some control for higher throughput and suits medium-risk workflows with reliable monitoring and a kill switch.

ISO 42001Governance

ISO/IEC 42001 is the international standard for an AI management system (AIMS). It gives organizations a certifiable framework of policies, roles, risk controls and continual improvement for developing and operating AI responsibly, including agentic systems.

LLM-as-a-judgeEvals

LLM-as-a-judge uses a language model to score or compare the outputs of an agent against criteria. It scales evaluation cheaply but must be calibrated against human labels, because judges show position, verbosity and self-enhancement biases.

MCP (Model Context Protocol)MCP & Identity

The Model Context Protocol (MCP) is an open standard for connecting AI models and agents to external tools, data sources and services through a uniform interface. Widely adopted in 2025, it also widened the attack surface around authorization and machine identity.

Memory poisoningSecurity

Memory poisoning corrupts an agent’s long-term memory, embeddings or retrieval store so that false or malicious content is recalled as trusted context later. Because the payload persists across sessions, it can quietly bias many future decisions.

Multi-agent systemOrchestration

A multi-agent system splits a task across several specialized agents that collaborate, often under a coordinator. It can improve modularity and quality but adds failure modes: agents may over-trust each other, amplifying errors and security risks across the system.

Non-Human Identity (NHI)MCP & Identity

A Non-Human Identity is the digital identity of a machine actor — an agent, service or workload — rather than a person. Agents authenticate with keys, tokens or certificates, so managing, scoping and rotating these identities is central to agent security.

ObservabilityObservability

Observability is the ability to understand an agent’s internal state from its outputs — traces, logs and metrics. For agents it means reconstructing each reasoning step, tool call and decision so failures can be diagnosed and behavior audited.

OrchestrationOrchestration

Orchestration coordinates how agents, tools and steps run together to complete a task. It covers control flow, state, retries and hand-offs, whether for a single agent or a multi-agent system, and frameworks like LangGraph or CrewAI implement different orchestration models.

Prompt injectionSecurity

Prompt injection is an attack that smuggles malicious instructions into the text an AI agent reads, overriding its intended behavior. Direct injection targets the user prompt; indirect injection hides instructions in documents, web pages or tool outputs the agent later processes.

Tool poisoningSecurity

Tool poisoning manipulates the tools an agent relies on — or their returned outputs — so the agent makes harmful decisions. It includes tampered tool descriptions, compromised APIs and crafted responses designed to steer the agent toward unsafe actions.

TracingObservability

Tracing records the full execution path of an agent run as a tree of spans: prompts, model calls, tool invocations and sub-agent steps, each with timing and metadata. OpenTelemetry’s GenAI conventions are becoming the de facto standard for this.

—