Core Vocabulary Every Agent Builder Should Know
Shared words prevent false agreement in design reviews.
//
Search across all documentation pages
Shared words prevent false agreement in design reviews.
This cheatsheet defines the terms you will see across agent loops, tools, memory, and orchestration - with a one-line "use it when" for each.
| Term | Definition | Use it when… |
|---|---|---|
| Agent | System that repeatedly decides among actions toward a goal under constraints | You need goal-directed multi-step behavior |
| Agent loop | Cycle of perceive → decide → act → observe until stop | Designing runtime control flow |
| Turn / step | One model decision cycle, possibly with tools | Budgeting max_turns |
| Trajectory / trace | Ordered record of messages, tool calls, and observations | Debugging and eval |
| Stopping condition | Rule that ends the loop (goal met, max turns, timeout, kill switch) | Preventing runaways |
| Autonomy level | How much the system may do without human steering | Scoping product risk |
| Human-in-the-loop (HITL) | Human approval or edit gates inside or around the loop | Irreversible or high-stakes actions |
| Initiative | Whether the system can act without a new user message | Proactive monitors vs reactive chat |
| Term | Definition | Use it when… |
|---|---|---|
| LLM / model | Generative model used as the decision/writing engine | Choosing capability vs cost |
| Context window | Max tokens of input+output the model can handle in one call | Designing history, tools, and RAG size |
| System prompt / policy | Durable instructions and constraints for the run | Encoding rules the model should follow |
| Messages / transcript | Conversational state sent each turn | Multi-turn agents |
| Tokens | Billing and length units for model I/O | Cost and truncation planning |
| Temperature / decoding | Sampling controls for randomness | Factual vs creative tasks |
| Structured output | Model response constrained to a schema (JSON, etc.) | APIs and reliable parsing |
| Reasoning trace | Intermediate chain-of-thought or provider "thinking" content | Debugging decisions (handle privacy) |
| Term | Definition | Use it when… |
|---|---|---|
| Tool / function calling | Model emits a named call + args; runtime executes it | Connecting models to systems |
| Tool schema | JSON schema describing name, args, types, description | Registration with the model API |
| Action space | Full set of tools and parameters the agent may use | Security and UX design |
| Observation | Tool result fed back into context | Closing the agent loop |
| Allowlist / permissions | Which tools are enabled for a run or role | Least privilege |
| Side effect | Externally visible change (send, write, charge) | Risk classification |
| Idempotency | Safe to retry without double effects | Unreliable networks and retries |
| Sandbox | Isolated environment for dangerous tools (shell, code) | Code-executing agents |
| Term | Definition | Use it when… |
|---|---|---|
| Grounding | Tying answers/actions to retrieved or tool evidence | Reducing pure hallucination |
| RAG | Retrieve documents, then generate with that context | Knowledge-base Q&A |
| Short-term memory | State within a run or session | Multi-step tasks |
| Long-term memory | Durable facts/preferences across sessions | Personalization, continuity |
| Working memory | What is actually in the context window now | Truncation and summarization |
| Entity memory | Structured store of people, tickets, objects | CRM-like agents |
| Citation | Pointer from claim to source chunk or tool result | Trust and audit |
| Term | Definition | Use it when… |
|---|---|---|
| Orchestration | Control layer that sequences models, tools, and humans | Multi-step systems |
| ReAct | Interleaved reasoning and acting pattern | Simple reactive agents |
| Plan-and-execute | Plan first, then run steps (sometimes replan) | Longer structured tasks |
| Multi-agent | Multiple specialized agents coordinating | Role split, parallel skills |
| Handoff | Transfer of control/context between agents | Specialist routing |
| Router | Component that picks model, agent, or tool path | Cost and skill specialization |
| Workflow / graph | Explicit state machine or DAG around model nodes | Controllable production flows |
| Framework | Library implementing loops, state, and integrations | Speed vs lock-in trade-offs |
| Term | Definition | Use it when… |
|---|---|---|
| MCP | Model Context Protocol - standard way to expose tools/resources to hosts | Sharing tools across apps/IDEs |
| A2A | Agent-to-agent communication patterns/protocols | Cross-system multi-agent work |
| Tool server | Process that hosts tools for one or more agents | Centralizing integrations |
| Provider gateway | Unified API across many model backends | Routing and failover |
| Term | Definition | Use it when… |
|---|---|---|
| Eval | Dataset + metrics for agent task success | Shipping changes safely |
| Faithfulness | Answer sticks to provided evidence | RAG quality |
| Prompt injection | Malicious instructions in user or tool content | Threat modeling |
| Guardrail | Filter/policy layer around inputs or outputs | Compliance and safety |
| Budget | Max turns, tokens, time, or dollars per run | Production cost control |
| Observability | Traces, metrics, logs for agent runs | Ops and debugging |
| Kill switch | Instant disable or autonomy downgrade | Incident response |
| Shadow mode | Agent proposes without executing side effects | Safe rollout |
| Loose phrase | Prefer |
|---|---|
| "The AI handles it" | "Level 3 agent, read-only tools, max 8 turns" |
| "We added memory" | "Session transcript + vector store for docs X" |
| "It's multi-agent" | "Router + researcher + writer with handoff schema Y" |
| "It's autonomous" | "Exception-only HITL; kill switch in flag Z" |
| "We grounded it" | "RAG over corpus C with citations; no answer if empty" |
No. Master agent loop, tool call, observation, stopping condition, context window, and trace first.
Mostly naming. Function calling is the common API phrase; tool use is the broader product phrase. Same idea.
State is whatever the runtime tracks for the current run. Memory often implies durable or retrieved knowledge beyond the raw transcript.
RAG is one grounding method. Tools (SQL, calculators, browsers) also ground actions in non-parametric evidence.
Rarely. Without action selection over an environment, you usually have chained generation, not an agent.
Turn index, model id, prompts or prompt hashes, tool name/args, observation summary, latency, token counts, stop reason.
Plan-and-execute drafts a multi-step plan up front; ReAct chooses the next action after each observation.
It is the set of real-world powers the agent has. Shrink it before you scale autonomy.
No. Tokens also bound how much history, tool output, and retrieved text fit in context.
This page plus a single traced demo run. Terms stick when tied to a real trajectory.
Make a one-page glossary in your repo: vendor name → term above → your implementation class/module.
Prompt is text to the model. Policy includes prompts plus hard runtime rules the model cannot override (tool denials, budgets).
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.
Reviewed by Chris St. John·Last updated Jul 16, 2026