Glossary — Agent Harnesses

§1

Core runtime

Agent: The running combination of a model, harness, capabilities, state, and current objective.
Agent harness: The control system around a model that assembles context, mediates tools, manages state, applies policy, and schedules inference.
Environment: The external reality in which observations and effects occur: files, processes, browsers, APIs, services, or people.
Inference: One runtime model computation that produces output from a particular input and sampling configuration.
Model: A trained system that maps an input sequence to a probability distribution over possible continuations.
Model adapter: A harness-owned translation boundary that converts internal messages, tool definitions, and model settings into one model-inference provider's request format, then normalizes streamed output, usage, and errors back into the harness vocabulary. Also called a model-provider adapter.
Message: A typed or role-tagged unit of input or output, such as system, user, assistant, tool, or a harness-specific event.
Prompt: The complete model input assembled for one inference, not merely the user’s latest sentence.
Step: One model inference plus the tool proposals and observations handled before the next inference.
Turn: The unit of work initiated by a user input and completed when the harness yields, stops, suspends, or requests more authority.

§2

Context and state

Compaction: A lossy transformation that replaces a larger history region with a smaller summary or structured digest.
Context: The ordered information made available to a model for one inference. In casual usage it may also refer to the budget or source material; specify which.
Context window: The maximum token capacity available to one model request and its generated output under the provider’s accounting rules.
Durable state: Information retained outside transient inference so it can survive later steps, restart, or retrieval.
Event stream: An ordered stream of typed lifecycle facts from which observers or state projections can derive views.
Memory: An umbrella term for retained information. Prefer transcript, working set, summary, artifact, retrieval index, profile, or durable state when that is what you mean.
Offloading: Moving content outside model-visible context while preserving a route to the retained data.
Prompt caching: Reusing computation for repeated model-input prefixes. It can change latency or cost without increasing logical context capacity.
Retrieval: Selecting information from a larger durable collection and copying a result into the current working set.
State: Information retained across runtime transitions: messages, events, settings, artifacts, summaries, approvals, or projections.
Transcript: A user-facing or event-level record of prior interaction, often larger and richer than the next model request.
Truncation: Keeping only a bounded portion of content and discarding the rest from that representation.
Working set: The selected and transformed information assembled for the next model inference.

§3

Tools and authority

Approval: A user or operator decision granting a bounded proposed action, scope, and duration.
Authority: The legitimate power delegated by a user or operator to perform a bounded action in an environment.
Capability: An operation the runtime can potentially perform. Possession of a capability does not establish permission to use it.
Confused deputy: A failure in which a more-privileged component is induced to use its authority for a requester that should not have it.
Permission: A policy decision about whether a particular principal may use a capability for a particular operation and target.
Policy: A rule set that decides whether, when, or under what constraints a capability may be used.
Prompt injection: An attempt to influence a model through untrusted content so it disregards or misapplies higher-authority instructions.
Sandbox: An execution boundary that limits resources available to a process even after an action is authorized.
Semantic guard: A validation rule that checks domain meaning beyond basic argument types, usually before side effects.
Tool: A named capability with a machine-readable input contract that the harness may expose to a model and execute.
Tool call: A model-proposed tool name plus arguments. It is a request for mediation, not proof of execution.
Tool result: A structured observation representing execution status, output, diagnostics, or a reference to larger retained data.

§4

Composition and delegation

Capability seam: A supported interface at which one provider of an operation can be selected or replaced without changing callers.
Extension: Code or resources loaded through a supported host contract to add or alter tools, events, instructions, UI, or lifecycle behavior.
Hook: A callback or command invoked at a documented lifecycle event, often to observe, transform, allow, or block behavior.
Plan: An explicit representation of intended work, dependencies, and progress used by a harness or user interface.
Plugin: A lifecycle-managed extension that registers behavior or services into a host composition system.
Projection: A derived view built from more durable source records, such as a transcript reconstructed from session events.
Service: A named runtime contract provided to other components through dependency or plugin composition.
Skill: A reusable package of instructions and optional resources that a harness can load for a task.
Subagent: A separately scheduled child agent run with its own context, task, tools, policy, and return channel.

← Comparison Sources →