Multi-Agent

Open standard for AI agent interoperability enabling seamless collaboration across platforms

Standardized communication mechanisms for agent interaction including message passing, publish-subscribe, and event-driven patterns

Orchestrator-worker architecture where a lead agent coordinates specialized subagents for parallel task execution

Multiple agents collaborate transparently on a common workspace with shared visibility of all work

Specialized agents process tasks in a linear pipeline where each agent's output becomes the next agent's input

Multiple agents work simultaneously on the same task to provide diverse perspectives and parallel processing power

Dynamic delegation where agents intelligently transfer control based on context and specialized capabilities

Decentralized collaboration between equal agents without central authority

Multi-level agent coordination with supervisory relationships, centralized orchestration, and authority delegation

Asynchronous message passing system enabling decoupled communication between AI agents and services in distributed agentic systems

Distributed agreement mechanisms for multi-agent decision making

Multiple AI models collaborate and debate to reach conclusions

Layered ensemble in which multiple LLM proposer agents at each layer receive all responses from the previous layer as auxiliary context, iteratively refining their answers before a final aggregator synthesizes them. Because models produce better answers when shown peer responses, a stack of open-source proposers can outperform a single frontier model.

An orchestrator drives a team of specialized agents using two ledgers: a task ledger holding facts, educated guesses, and a plan, and a progress ledger where the orchestrator self-reflects each step on whether progress is being made, whether the team is stalled, and which agent should act next. When progress stalls, an outer loop updates the facts and revises the plan. Introduced in Microsoft Magentic-One.

A training-free framework for long-context tasks that splits an oversized input into chunks, assigns each chunk to a worker agent that reads its chunk plus a running communication unit passed from the previous worker, then hands the final accumulated communication to a manager agent that produces the answer. The sequential read-and-summarize chain mitigates long-context focus loss and beats RAG and full-context baselines.