When running a multi-agent system, we need a way for agents to communicate with one another. There are quite a few ways for the agents to talk/pass info to each other.
Linear
This is when the flow is set to a series of steps, one after another, just a simple workflow like drafting replies to a set of emails or generating captions for images.
Hierarchical
When we have an agent or an agent harness that orchestrates multiple sub-agents to get a job done. These orchestrators can be deterministic code or another agent.
Deterministic Orchestrator
- Generate-and-Critic Pattern: Your code spawns one model to generate text and an adversarial model to poke holes and give feedback.
- Planner Agent: Implements the generate-and-critic pattern where one model generates a plan to achieve the goal, hands it to the orchestrator, and the orchestrator asks the adversarial model to identify issues with the plan.
Agent as Orchestrator
- Decomposition: A pattern in which a single query is split into multiple subqueries, each handled in parallel. For instance, if the query requires facts from multiple sources (web search, database lookup, internal docs). Agentic RAG is an example of such a pattern.
Networks or Graph
When we have a system with multiple agents communicating, all are free to communicate with each other without depending on a central harness. These are for fully autonomous agents where they can spawn other agents and work together to achieve a task.
Council Pattern
- Multiple Experts: A single query is sent to multiple specialised agents (e.g., a security expert, a performance expert, and a UX expert).
- Consensus or Synthesis: The agents share their findings to reach a consensus, or a final "moderator" agent synthesises their independent responses into a single, comprehensive output.
Multi-Agent Debate
- Adversarial Collaboration: Two or more agents are given a task and are encouraged to critique each other's work over multiple rounds.
- Self-Correction: This iterative peer-review process within the network helps eliminate hallucinations and improve the technical accuracy of the final result.