Glossary
56 canonical definitions. One term, one meaning, everywhere on this site.
Adoption scoreAgentiquette's log-scaled, 5%-capped measure of community traction (stars, forks, dependents). Deliberately the weakest …Agent instruction fileThe umbrella term for persistent, project- or tool-scoped files that configure agent behavior: AGENTS.md, CLAUDE.md, Cur…Agent loopThe control flow wrapped around an agent's skills and tools that forces work to completion: plan, act, check the result,…Agent memoryAny mechanism that persists information across agent sessions so experience converts into improvement: memory files, vec…Agent skillA named, triggerable procedure that teaches an agent how to do a specific class of work: purpose, trigger conditions, st…Agentic execution systemThe reusable infrastructure that makes an AI agent reliable: the skills that encode how work is done, the loops that for…AGENTS.mdA persistent instruction file at a repository's root that tells any AI agent how to work in that project: setup commands…Audit log (agents)An append-only record of agent actions written by infrastructure the agent does not control, providing ground truth abou…BoundednessThe property that a loop cannot run away: budgets on iterations, cost, and time, with defined behavior (abort, escalate)…Budgeted retryRetrying failed operations under an explicit budget with a required diagnosis between attempts, escalating with evidence…Canonical definitionA single, byte-identical definition of a term, maintained at one stable URL and repeated verbatim wherever the term appe…Checkpoint-resumeA loop pattern for work that outlives a session: progress is persisted to a state file at durable checkpoints, and every…CLAUDE.mdClaude Code's project-scoped instruction file, loaded automatically at session start: project context, commands, code st…Closed-loop agent workflowA workflow whose output is checked against its goal by the workflow itself, with failures routed back for correction, in…CompactionCompressing a session's context when it approaches the window limit, typically by summarizing earlier content so work ca…Context windowThe bounded working memory of a model session: everything the agent currently knows, including instructions, conversatio…Cursor rulesInstruction files for the Cursor editor (.cursor/rules or .cursorrules) that apply persistent conventions and triggered …Episodic memoryAgent memory that records what happened: session events, decisions taken, outcomes observed, usually as logs or session …Eval setA fixed collection of test cases with expected properties, run repeatedly against an agent or prompt to measure quality …EvaluationProving an agent's work is correct, repeatedly and measurably: test suites, benchmarks, rubrics, and verification checkl…Evaluation readinessThe degree to which an artifact's success can be measured: does it produce checkable outputs, define success criteria, s…False successAn agent declaring work complete or correct when it is not: the most expensive agent failure mode, because it converts a…Fresh-context reviewReview performed by an agent (or session) that has not seen the producer's reasoning, so it cannot inherit the misunders…Governance (agents)The layer that keeps agents safe, scoped, and accountable: permission systems, boundaries inside skills, audit logs, and…GuardrailsProgrammatic checks on agent inputs and outputs that enforce policy: validation, filtering, correction, or refusal when …HandoffThe transfer of a task from one agent to another, including the context the receiving agent needs and the criteria for w…Hook (agent lifecycle)Code that runs automatically at defined points in an agent's lifecycle (session start, tool call, completion), outside t…Human gateA required human approval point inside an otherwise autonomous workflow, placed where an action's reversal cost exceeds …LLM-as-judgeUsing a model to evaluate another model's output against a rubric, standing in for human judgment on tasks without mecha…Loop reliabilityThe quality of a loop pattern measured on five axes: completion rate, verification quality, recoverability, observabilit…Memory poisoningA wrong or malicious write to agent memory that later recalls as trusted fact, corrupting every downstream decision that…Memory safety (agents)The property that an agent's memory helps rather than harms: relevant on recall, fresh or verifiably dated, private wher…Model Context Protocol (MCP)An open protocol standardizing how AI agents connect to tools and data sources: servers expose capabilities with typed s…Observability (agents)The ability to see what an agent system actually did: traces of steps, tool calls, decisions, and costs, sufficient to e…Open-loop executionFire-and-forget agent work: output is produced and accepted without the workflow checking it against the goal. The oppos…OrchestrationCoordinating multiple agents, tools, and steps into a workflow: who does what, in what order, with what handoffs, under …Permission profileThe declared set of actions an agent may take without asking: which tools, which paths, which commands, and what require…Plan-execute-verify (PEV)A loop pattern in which the agent writes an explicit plan with per-step success criteria, executes one step at a time, a…Procedural memoryAgent memory that stores how to do things. In practice, skills are externalized procedural memory: procedures made durab…Prompt injectionAn attack where content an agent processes (a web page, a README, an email) carries instructions that hijack the agent's…Red-teaming (agents)Systematically attacking your own agent system (injection, jailbreaks, data exfiltration, unsafe tool use) to find failu…Reviewer loopA loop pattern in which work passes through a fresh-context review against a written rubric before acceptance, with find…Role prompt (persona)An identity-layer instruction defining what an agent is (a security reviewer, a staff engineer) to shape its expertise a…Semantic memoryAgent memory that stores facts independent of when they were learned: project conventions, user preferences, environment…Skill leverageHow much reliable capability one artifact adds per unit of adoption effort. High-leverage skills change many tasks' outc…Skill marketplaceA distribution system for skills, commands, and agent components: install tooling plus a catalog, trading curation depth…SKILL.mdA structured instruction file used to teach AI agents a reusable capability, workflow, or domain behavior. A strong SKIL…Slash commandA manually-invoked skill: a named procedure the operator triggers explicitly (/ship, /review) rather than one the agent …StalenessThe gap between what a stored fact says and what is currently true. The dominant failure mode of agent memory and frozen…SubagentAn agent spawned by another agent to handle a scoped task with its own fresh context, returning results to the orchestra…Task decompositionBreaking a goal into ordered, checkable subtasks, each small enough to verify individually. The plan phase of every seri…Tool useThe mechanism by which an agent acts on the world: invoking functions, APIs, shells, and browsers, then observing the re…TracingRecording the full execution tree of an agent run (prompts, tool calls, intermediate outputs, timings, costs) as structu…Trigger conditionsThe part of a skill that states when it applies: the task shapes, phrases, or situations that should cause the agent to …Verification stepA check that evaluates captured evidence (command output, file state, rendered result) against a pre-stated criterion, a…Working memory (scratchpad)Task-scoped memory for the current job: plans, intermediate results, open questions. Deliberately ephemeral; it should n…