Multi-Agent Glossary

Agents and workflows

Agent

An LLM-driven system that can decide steps, call tools, inspect results, and continue toward a goal. The key distinction is not personality or naming. It is whether the model controls enough of the loop to choose what happens next.

Workflow

A system where LLM calls and tools follow a predefined path. Prompt chaining, routing, parallel checks, and evaluator loops are workflows when the control path is coded by the developer rather than chosen freely by the model.

Single-agent system

One agent owns the loop and uses tools, retrieval, memory, and guardrails to complete the task. It is often the right baseline because context is continuous and evaluation is simpler.

Multi-agent system

Multiple agents coordinate on one user goal. They may be arranged under a manager, connected by handoffs, run in parallel, debate, or operate through shared artifacts. The defining issue is coordination, not the number of model calls.

Pattern terms

Orchestrator

The lead agent that decomposes a task, delegates work, receives artifacts, and synthesizes the final answer. A good orchestrator owns acceptance criteria and source quality, not merely scheduling.

Worker

A specialized agent assigned a bounded subtask. A worker should return a verifiable artifact: source notes, tests, code changes, classifications, or structured results.

Pipeline

A fixed sequence of steps. It may use named agents, but the order is not dynamically invented. Pipelines are useful when the task can be decomposed reliably before runtime.

Debate

A pattern where multiple agents or model instances propose answers, critique one another, and converge through a judge, vote, or final synthesis.

Swarm

A loosely coordinated group of agents. The term is often vague; a serious design should specify message rules, shared memory, stop conditions, and who is accountable for the final answer.

Coordination terms

Shared memory

Durable state accessible across agents. Good shared memory is artifact-based: claims, sources, plans, decisions, tests, and open questions. It is not a dump of every token.

Message passing

The protocol agents use to request work, return results, and ask for more context. Strong messages include boundaries, expected artifacts, and stop conditions.

Handoff

A transfer of control from one agent to another. Handoffs are useful when the next role has different tools, policy, or domain ownership; they are dangerous when they omit the relevant decisions that got the task to its current state.

Context isolation

Giving agents separate context windows so each can focus on a subproblem. Isolation can increase capacity, but it must be paired with artifacts so important decisions do not disappear between agents.

Risk terms

Tool overload

A failure mode where one agent has too many similar tools and chooses poorly. Splitting agents can help only if each new role has clearer tool ownership.

Context fragmentation

Loss of task coherence because different agents see different context and make incompatible implicit decisions.

Error compounding

Downstream failures caused by upstream mistakes being treated as valid state. Multi-agent systems need checkpoints and artifact validation to reduce this.

Evaluator-optimizer

A workflow where one model generates output and another evaluates it in a loop. It is valuable when evaluation criteria are clear and iterative critique measurably improves the result.