WaggleDance -- Multi-Agent Communication

Single-User Multi-Agent

One user, multiple specialist agents working in parallel. The SubagentOrchestrator resolves dependencies, injects context, and synthesizes results into a unified answer.

👤

User

🎯

Orchestrator

🔍

Researcher

💻

Coder

📊

Analyst

1. Decompose

2. Spawn

3. Execute

✔

Merged Result

4. Return

👤

User

Team WaggleDance

Multiple users share a team memory layer via TeamSync. Knowledge discovered by one agent flows to all team members automatically -- push-on-write, pull-on-activate.

👤

User A

🐝

Agent A

Discovers

🧠

Local Mind A

Writes

🏠

Shared Team Memory

Push

Pull

🧠

Local Mind B

🐝

Agent B

👤

User B

Recalls

TeamSync: frameToEntity() -- composite key gopId+t

👤

User C (also synced)

Task Dispatcher

The WaggleDanceDispatcher routes messages by type/subtype combination. task_delegation spawns workers, knowledge_check searches memory, skill_request resolves capabilities.

📋

Task Delegation

❓

Knowledge Check

🔧

Skill Request

🎯

Dispatcher

🔎

Cap. Match

⚖

Worker Spawn

🧠

Memory Search

🐝

Agent Pool

🐝

New Worker

📦

Results

Protocol: request=[knowledge_check, task_delegation, skill_request] | response=[knowledge_match, task_claim] | broadcast=[discovery, routed_share, skill_share]

HiveQuery Broadcast

One agent broadcasts a question to the entire hive. Every agent searches its knowledge, and results are collected, ranked by relevance, and the best answer surfaces. Like a hive mind where every bee contributes.

📢

Broadcast Query

HiveQuery { topic, scope }

🐝

Researcher

🐝

Coder

🐝

Analyst

🐝

Writer

🐝

Reviewer

🏆

Collect + Rank

⭐

Best Answer

Surfaces

Shared Resources

What flows between agents in the WaggleDance protocol. Every shared resource type maps to a real mechanism in the codebase.

🧠

Memory Frames

Shared via team-sync.ts. Push-on-write, pull-on-activate. Frames carry author attribution (userId + displayName).

team-sync.ts

🎯

Skills

Shared via skill_share broadcast message. Receiving agents install shared skills into their workspace skill library.

dispatcher.ts

🕸

Knowledge Graph

Entity-relation graph with bitemporal validity. Global KG queries span all workspaces via MultiMindCache.

knowledge.ts

🧬

Learnings

Evolution improvements propagate: persona overrides, behavioral-spec patches. Accepted runs improve all agents globally.

evolution-deploy.ts

🛠

Tool Pool

Role-based tool presets shared across agents. 8 roles with curated tool sets. filterToolsForContext() enforces access.

subagent-orchestrator.ts

Communication Protocol

The lifecycle of a sub-agent from spawn request to trace recording. Each step maps directly to code in the agent package.

1

Spawn Request

Orchestrator receives a WorkflowStep with role, task, tools, and dependency list. Validates that all dependsOn steps have completed before proceeding.

2

Agent Creation

Worker ID generated via makeWorkerId(). WorkerState initialized as pending. The worker:status event fires so UI can show the planned workflow.

3

Context Injection

setParentContext() injects user identity, workspace info, and behavioral rules. Results from contextFrom steps are prepended as prior context. Role tool presets resolved from ROLE_TOOL_PRESETS.

4

Execution

Status transitions: pending -> running -> done|failed. Agent loop runs with constrained tools and turn budget. Token usage tracked in WorkerState.usage.

5

Result Return

Worker result stored in WorkerState.result. Context map updated for downstream steps. Final aggregation mode: concatenate | last | synthesize.

6

Trace Recording

TraceRecorder captures full execution trace: tokens, tools, latency, model, quality. Traces feed into the evolution pipeline -- enabling closed-loop self-improvement.