agent-conversation-protocols

license	Apache-2.0
name	agent-conversation-protocols
description	Conversation patterns and interaction protocols for multi-agent systems. Covers request/response, pub/sub, blackboard, delegation chains, debate, critique, consensus, fan-out/fan-in, supervisor-worker, and peer negotiation. Deep analysis of AutoGen conversation patterns, CrewAI delegation, LangGraph state passing, and FIPA-ACL performatives. Teaches how to design what agents say to each other and in what order. Activate on: "agent conversation", "agent protocol", "multi-agent debate", "agent delegation", "supervisor worker pattern", "agent voting", "consensus protocol", "fan-out fan-in", "agent negotiation", "blackboard pattern", "agent dialogue", "conversation topology", "agent handoff". NOT for: wire format or serialization (use agent-interchange-formats), orchestration infrastructure (use agentic-infrastructure-2026), single agent behavior (use agentic-patterns).
allowed-tools	["Read","Write","Edit","Bash","Glob","Grep","WebSearch","WebFetch"]
metadata	{"category":"AI & Agents","tags":["agents","protocols","conversation","delegation","debate","consensus","multi-agent","orchestration"],"pairs-with":[{"skill":"agent-interchange-formats","reason":"Interchange formats define the wire; this skill defines the dialogue"},{"skill":"multi-agent-coordination","reason":"Coordination is the topology; conversation is what flows through it"},{"skill":"agentic-infrastructure-2026","reason":"Framework choice constrains available conversation patterns"},{"skill":"agentic-patterns","reason":"Single-agent patterns compose into multi-agent conversations"}]}
category	Agent & Orchestration
tags	["agent","conversation","protocols","ai","orchestration","design"]

Agent Conversation Protocols

You are an expert in multi-agent conversation design. You understand how agents talk to each other -- the message types, turn-taking rules, delegation patterns, and conflict resolution mechanisms that make multi-agent systems coherent rather than chaotic.

DECISION POINTS

Primary Pattern Selection Tree

Given problem characteristics:

├── Task is decomposable into independent subtasks?
│   ├── YES + Quality matters more than speed
│   │   └── Use FAN-OUT/FAN-IN with redundant execution (3+ agents same task)
│   ├── YES + Speed matters more than quality  
│   │   └── Use FAN-OUT/FAN-IN with partitioned execution (divide work)
│   └── NO + Task requires sequential dependencies
│       └── Use SUPERVISOR-WORKER with delegation chains
│
├── Multiple valid approaches exist?
│   ├── YES + Verifiable ground truth exists
│   │   └── Use DEBATE (adversarial refinement with judge)
│   ├── YES + Subjective preference decision
│   │   └── Use VOTING/CONSENSUS (democratic selection)
│   └── NO + Single approach but needs refinement
│       └── Use CRITIQUE-REFINE (iterative improvement)
│
├── Knowledge synthesis from diverse sources?
│   └── Use BLACKBOARD (shared state accumulation)
│
└── Simple capability delegation?
    └── Use REQUEST/RESPONSE (synchronous handoff)

Topology × Initiative × Turn Order Decision Matrix

Topology	Initiative	Turn Order	Use When	Example
Star	Push	Round-robin	Clear leader coordinates work	CrewAI hierarchical process
Star	Pull	Priority-queue	Workers request tasks when ready	AutoGen GroupChat with manager
Mesh	Push	Free-form	Peer collaboration, no bottlenecks	Multi-agent debate
Tree	Push	Depth-first	Hierarchical decomposition	Complex delegation chains
Broadcast	Reactive	Event-driven	Knowledge sharing, updates	LangGraph state updates

Termination Condition Selection

If conversation type is:
├── DEBATE → Stop when judge_confidence > 0.8 OR rounds >= 3
├── CRITIQUE → Stop when verdict == 'approve' OR iterations >= 4  
├── VOTING → Stop when all votes collected OR timeout
├── FAN-OUT → Stop when gather_policy satisfied (all/majority/first)
├── SUPERVISOR → Stop when all subtasks complete OR budget exceeded
└── BLACKBOARD → Stop when goal_condition met OR staleness detected