Glossary

Term	Meaning
Agent	A system that pursues a goal, invokes permitted tools, observes results, and iterates within a harness.
Ask / Plan / Agent	Roles used to investigate, design, and execute. They are not environments or models.
Context	Information available to the current task, including conversation, files, instructions, selections, and tool results.
Custom agent	A `.agent.md` role definition with instructions, tools, and optional handoffs.
Environment	Where files and processes exist: folder, worktree, Codespace, local machine, or remote ephemeral workspace.
Evidence	Observable support for a claim: command output, exit status, test result, diff, trace, or review record.
Harness	Runtime that hosts an agent and mediates tools, permissions, context, sessions, and lifecycle.
Handoff	Transfer of a concrete task or artifact between roles, agents, harnesses, or environments.
Hook	Deterministic command invoked at a documented lifecycle event. Hooks are Preview in VS Code as of 2026-09-05.
Instructions	Guidance applied automatically, globally or for matching paths or tasks.
Model Context Protocol (MCP)	Protocol for connecting AI hosts to external tools, resources, and prompts.
Prompt file	Manually invoked reusable task template, primarily associated with the VS Code Local harness.
Session	Conversation, context, tool calls, and execution state managed by an agent harness.
Skill	Reusable expertise and resources loaded when relevant to a task.
Subagent	Focused agent invoked in separate context that returns a result to a coordinating agent.
Target	Destination selected for a session, such as Local, Copilot, or Cloud in VS Code.
Tool	Permitted operation such as reading, searching, editing, running a command, or calling an external system.
Worktree	Git working tree for a separate branch. It isolates code changes, not security permissions.

[!NOTE] Model, role, harness, target, and environment answer different questions. Avoid using “agent” as a catch-all for all five.