Glossary / 2026
AI Agent Topology Glossary (2026)
A focused glossary of topology-specific terms. Definitional terms (what is an AI agent, types of agents, agent loop) live on whatisanaiagent.com’s glossary.
- Augmented LLM#augmented-llm
- A language model that can retrieve information, call tools, and remember context across a turn-by-turn loop. The foundational unit in Anthropic's Building Effective Agents framing.
- Centralised orchestration#centralised-orchestration
- An agent topology where a single agent (the orchestrator or supervisor) routes work to other agents and aggregates results. Synonym for the supervisor pattern.
- Coordinator pattern#coordinator-pattern
- Synonym for supervisor or orchestrator pattern, used in some framework documentation. The coordinator decides who speaks next or who acts next.
- Crew (CrewAI)#crew
- A CrewAI primitive describing a group of agents with assigned roles working together on a task. The canonical CrewAI shape is a sequential crew where tasks pass through researcher, writer, and reviewer roles.
- Emergent orchestration#emergent-orchestration
- An agent topology where coordination happens via direct peer-to-peer messaging rather than via a central supervisor. Closer to a peer multi-agent or swarm pattern.
- Evaluator-optimiser pattern#evaluator-optimiser
- A two-agent loop in which a generator produces a candidate output and an evaluator scores it against criteria. The generator revises until criteria are met or iterations exhaust.
- Generator-critic#generator-critic
- An informal name for the evaluator-optimiser pattern. The generator produces; the critic evaluates; the loop iterates.
- Group chat (AutoGen)#group-chat
- An AutoGen primitive in which multiple agents share a conversation history and a GroupChatManager selects which agent speaks next. The shape is a peer multi-agent topology with a coordinating manager.
- Handoff#handoff
- A message-passing primitive in which one agent transfers control to another, optionally carrying context. Common in LangGraph (handoff messages between supervisor and workers) and OpenAI's Agents SDK.
- Hierarchical agents#hierarchical-agents
- An agent topology with multiple levels of supervision: a top supervisor delegates to team-lead supervisors, who themselves delegate to workers.
- Human-as-arbiter#human-as-arbiter
- A human-in-the-loop variant where the agent operates autonomously on routine cases and escalates ambiguous or sensitive cases to a human. Klarna's customer service deployment is a canonical example.
- Human-as-reviewer#human-as-reviewer
- A human-in-the-loop variant where the agent does the bulk of the work autonomously, then pauses for human approval before any irreversible action. LangGraph's interrupt() is the canonical implementation.
- Human-in-the-loop#human-in-the-loop
- An agent topology in which a human role reviews or approves agent actions before they take effect. Three named variants: assistant, reviewer, arbiter.
- Manager agent#manager-agent
- An agent at a higher level of a hierarchy whose role is to delegate to other agents and aggregate their outputs. Synonym, in this context, for a top-level supervisor or orchestrator.
- Multi-agent system#multi-agent-system
- Any agent topology with two or more agents that interact. Subdivides into peer multi-agent (no central supervisor), supervisor patterns, hierarchical patterns, and human-in-the-loop variants.
- Orchestrator-workers#orchestrator-workers
- Anthropic's name for the dynamic-spawn variant of the supervisor pattern: the orchestrator decomposes a goal at runtime and spawns worker agents per sub-task.
- Peer agents#peer-agents
- Agents that interact without a central supervisor, exchanging messages via a shared bus or by direct handoff. CrewAI sequential crews and AutoGen group-chats are common implementations.
- Routing (static / LLM-judged / tool-based)#routing
- How a supervisor selects the next worker. Static: a deterministic rule. LLM-judged: the supervisor's LLM picks. Tool-based: workers are exposed as callable tools and the supervisor's tool-call selects the worker.
- Single-agent topology#single-agent-topology
- An agent topology with one agent running a sense-think-act loop with access to a model and a set of tools. The simplest case and the most common in production.
- Supervisor agent#supervisor-agent
- An agent whose role is to decompose goals and route sub-tasks to specialised worker agents, then aggregate the results. The defining role in the supervisor pattern.
- Swarm pattern#swarm-pattern
- A peer multi-agent shape in which agents hand control to each other directly via handoff messages, with shared state and no central router. Documented in LangGraph as the swarm primitive.
- Worker agent#worker-agent
- An agent whose role is to execute a specialised sub-task within a supervisor or hierarchical pattern. Workers are typically specialised by tool, role, or domain.
Source: Anthropic, Building Effective Agents (Schluntz, December 2024).
Read more: /single-agent/
Read more: /supervisor-pattern/
Source: CrewAI documentation, accessed 30 April 2026.
Read more: /multi-agent/
Read more: /multi-agent/
Source: Anthropic, Building Effective Agents (Schluntz, December 2024).
Read more: /evaluator-optimiser/
Read more: /evaluator-optimiser/
Source: AutoGen documentation, accessed 30 April 2026.
Read more: /multi-agent/
Read more: /supervisor-pattern/
Source: LangGraph documentation, accessed 30 April 2026.
Read more: /hierarchical-agents/
Read more: /human-in-the-loop/
Read more: /human-in-the-loop/
Read more: /human-in-the-loop/
Read more: /hierarchical-agents/
Read more: /multi-agent/
Source: Anthropic, Building Effective Agents (Schluntz, December 2024).
Read more: /supervisor-pattern/
Read more: /multi-agent/
Read more: /supervisor-pattern/
Read more: /single-agent/
Source: Anthropic, Building Effective Agents (Schluntz, December 2024); LangGraph documentation.
Read more: /supervisor-pattern/
Source: LangGraph documentation, accessed 30 April 2026.
Read more: /multi-agent/
Read more: /supervisor-pattern/
For the broader AI agent vocabulary (definition, types, agent loop, evaluation), see whatisanaiagent.com/ai-agent-glossary. For the engineering reference, see buildingeffectiveagents.com.