A multi-agent system is an architecture in which two or more autonomous AI agents operate within a shared environment, each perceiving inputs and taking actions that may affect the others. Agents can be cooperative - dividing subtasks to reach a shared goal - or adversarial, as in the generator-discriminator pairing inside a GAN. Modern LLM-based multi-agent frameworks extend this concept to language models that delegate, critique, or verify each other's outputs through structured message passing.

Think of it like a law firm where a partner drafts a contract, an associate reviews it for errors, and a paralegal cross-checks cited statutes - each role is specialised, but together they produce a result none could deliver as quickly alone.

A concrete implementation might use a planner agent and an executor agent:

# LangGraph-style two-agent loop (pseudocode)
planner = Agent(llm=gpt4, role="Break task into steps")
executor = Agent(llm=gpt4, role="Execute one step and report result")

plan = planner.run(user_task)
for step in plan.steps:
    result = executor.run(step)
    planner.update(result)       # planner can revise remaining steps

Related ideas include Prompt Engineering for crafting inter-agent instructions, Chain of Thought reasoning that individual agents use internally, RAG for giving agents access to external knowledge, and RLHF for fine-tuning agents that learn from feedback. The Transformer backbone powering each LLM agent also ties directly into the broader Neural Network foundation.

You are most likely to encounter multi-agent systems in autonomous research pipelines, software-engineering copilots, and agentic workflow products where a single model call is insufficient for a complex, multi-step task.