Glossary of AI agent terms

An autonomous agent is a software system that pursues a goal across many steps and channels without a person driving each action. It decides what to do next from the current state of the work, acts, observes the result, and continues until the goal is met or it escalates.

An agent runtime is the harness that keeps a long-running agent coherent across time — storing its memory, scheduling its future work, and recording every action. It is the difference between an agent that survives a week-long gap and one that forgets the job when a session ends.

A long-running workflow is a unit of work that spans days or weeks, with gaps, external waits, and many touches before it completes. Unlike a single request, it only counts as finished when a real-world outcome has changed, such as a bound policy or a posted payment.

Orchestration is the coordination of multiple systems, steps, and waits toward a single outcome. In an agent, orchestration means sequencing calls, approvals, and provisioning across tools that do not talk to each other, and re-sequencing when a dependency slips.

Human-in-the-loop is a design where a person approves or reviews specific agent actions before or after they happen. It keeps judgment and authorization with a human for high-stakes steps while the agent handles the routing, chasing, and execution around them.

Human escalation is the act of handing a job to a person when the agent reaches a limit it should not cross alone. A well-designed escalation includes the full history and context, so the person resumes the work rather than restarting it.

Agent memory is the durable record of everything that has happened on a job, persisted so it survives restarts, deploys, and long gaps between touches. It is what lets an agent read its own earlier calls and decisions weeks later instead of starting fresh each session.

A policy engine is the component that decides an agent’s next action from the accumulated state of the job. It separates judgment from execution, so decisions are made by inspectable rules rather than by whatever a model happens to output on a given run.

Deterministic decisioning means the same facts always produce the same decision. In regulated work it is what makes an agent auditable: a compliance team can replay why an action was taken instead of accepting that a model simply felt like it.

A multi-channel agent works across voice, SMS, email, portals, and APIs within a single job, rather than being confined to one surface. It chooses the channel each step needs and carries the same memory of the job across all of them.

A voice agent makes and receives phone calls as part of its work — navigating phone trees, waiting on hold, and speaking with a person. Every call is transcribed and fed back into the job as the record of what happened.

Outcome-based pricing charges for jobs that close rather than for activity like minutes, messages, or model calls. It aligns the vendor with the customer’s goal — the bound renewal or paid claim — because effort that does not reach an outcome is not billable.

Ticket deflection is resolving a request without a person opening a ticket queue at all. An agent deflects a ticket by owning the request end to end — approval, provisioning, and verification — rather than just answering a question and leaving the work to someone else.

Straight-through processing is completing a transaction end to end without manual intervention. For an agent it means carrying a job from trigger to closed outcome — a bound policy, a reprocessed claim — with a person involved only at the decision points that require one.