Multi-Agent Systems

Key jargon

Term	Plain-language meaning
Multi-agent system	An orchestration using multiple independently prompted or specialized agent roles.
Delegation	Assigning a bounded responsibility and expected result to another worker.
Coordinator	The component that decomposes work, resolves dependencies, and integrates results.
Consensus	A method for selecting or reconciling outputs from multiple participants.

Key concepts

Multiple agents add communication and integration failure modes; they are useful only when decomposition or diversity pays for that cost.
Independent evidence or parallel work is more valuable than several agents repeating the same reasoning.

Concept map

flowchart LR
    A["Coordinator decomposes goal"] --> B["Agents work on bounded tasks"]
    B --> C["Validate evidence and conflicts"]
    C --> D["Integrate one result"]

Multiple agents do not automatically create better reasoning. They can add parallelism and specialization, but also duplicate mistakes, amplify unsupported consensus, and consume more context and budget.

When separation helps

Independent subtasks can run concurrently.
Different tools, data permissions, or expertise are required.
An adversarial or evidence-focused reviewer is genuinely independent.
Artifacts have clear ownership and deterministic merge rules.

Coordination contract

Define task, inputs, allowed tools, owned artifacts, budget, output schema, evidence, completion condition, and merge authority for each worker. Use shared durable state rather than relying on agents to remember conversation.

Exercise

Split a product comparison among researcher, cost analyst, and verifier. Define what each owns and how conflicting facts are reconciled. Then compare with one agent using three sequential roles.

Misconception

Several agents agreeing is not independent corroboration when they share the same model, prompt, context, or source error.