Curriculum

Agentic AI · Collaborative Agents

Multi-Agent Systems

Multi-agent systems coordinate fleets of specialised AI agents, each with distinct roles, tools, and memory. Hierarchical orchestration and role-based collaboration enable parallelism and specialisation.

The idea

A single LLM agent has a finite context window and limited specialisation. Multi-agent architectures distribute tasks across a network of agents: an orchestrator decomposes the goal and delegates subtasks to worker agents. Workers report results back, the orchestrator synthesises and decides next steps.

Role-based systems (CrewAI, AutoGen) assign each agent a persona, backstory, and tool set. A researcher agent uses web search; a coder agent executes Python; a reviewer agent checks outputs. Communication between agents is structured as messages in a shared thread or direct API calls.

Hierarchical Task Decomposition: the orchestrator uses an LLM call to break the high-level goal into a dependency-ordered task tree, then schedules tasks respecting dependencies. Parallel branches are executed concurrently for speed.

T = {t₁, t₂, …, tₙ} tᵢ → agentᵢ(tools, memory)

Task set T distributed to specialised agents. Each agent operates with its own tool set and memory.

What to hold onto

  • Consensus MechanismsMultiple agents vote on decisions (majority, Borda count, or weighted by track record) to reduce single-agent failure modes.
  • Agent Memory SharingShared vector stores allow agents to write and retrieve knowledge, enabling long-running collaborative workflows across sessions.
  • Failure ModesMiscommunication between agents, circular delegation, undetected errors cascading downstream. Observability and structured handoffs are essential.

Live: Hierarchical Multi-Agent Consensus Protocol

Manager Agent delegates sub-tasks to Researcher, Coder, and Critic agents with message-passing synchronization.

Consensus: Verified

Quick answers

What are Multi-Agent Systems in AI?

Multi-agent systems coordinate multiple specialised AI agents to solve complex tasks collaboratively. An orchestrator agent decomposes goals and delegates subtasks to specialist workers with distinct tools and roles. Agents communicate via structured messages, enabling parallelism, specialisation, and resilience beyond single-agent limits.

What is hierarchical task decomposition in multi-agent systems?

Hierarchical task decomposition breaks a complex goal into a dependency-ordered tree of subtasks. The orchestrator assigns leaf tasks to specialist agents, which execute in parallel where possible. Results are progressively synthesised up the tree until the root task is complete.