Cfn Coordination

Skill: cfn-coordination

Purpose

Patterns for coordinating multiple Task-mode agents within a single main-chat message. Replaces the deprecated Redis pub/sub coordination model.

Architecture Principle

Task-mode agents return output directly. No Redis signaling. Main chat (or a coordinator agent) reads each agent's return value and routes data to dependents. Agents never block on external coordination channels.

Anti-patterns (do not use):

• redis-cli lpush / blpop between agents
• Pub/sub swarm:* channels
• Agents subscribing to feedback queues mid-execution

Core Patterns

1. Chain (sequential dependency: A → B → C)

Each step waits for the prior step's return value. Main chat (or coordinator) feeds output forward.

• Use when later stages need exact output of earlier stages
• Each link must be a separate spawn round; you cannot truly chain inside one message
• For 2-stage chains within one batch, the coordinator agent handles the second stage after reading the first agent's return

2. Broadcast (1:N: A → B, C, D)

One agent produces output; multiple agents receive the same input.

• Coordinator reads source agent's return value once, then includes it in each dependent agent's prompt
• All N dependents can spawn in the same batch after the source agent completes

3. Mesh (N:1 or N:N: A, B, C → D)

Multiple agents produce output in parallel; one or more downstream agents consume the aggregate.

• Spawn the N producers in a single message
• After all return, coordinator (or main chat) aggregates and spawns the consumer
• For pure parallel work with no consumer, just spawn the batch and read all returns

4. Consensus Collection (validators voting)

Multiple validator agents review the same artifact and return confidence scores. Used in Loop 2.

• Spawn validators in a single message with identical artifacts as input
• Aggregate confidence scores (mean, or per-rubric)
• Pass/fail against a threshold (e.g., 0.90 consensus for Loop 2)

Topology Selection

Task-Type Cookbook

Agent Prompt Template

Task("<agent-role>", `
  <task description>

  Inputs:
  - <data piped in from prior agents, inlined into prompt>

  Deliverables:
  - <files / artifacts / decisions to produce>

  Return format:
  - Final message must include confidence score (0.0-1.0) and list of deliverable paths
`, "<agent-type>")

Notes:

• No Redis channels, no blocking waits. Output is the return value.
• Confidence score is required for Loop 2/3 gate checks. See cfn-loop-orchestration-v2 skill.
• Deliverable paths must be absolute or repo-relative, no glob patterns.

Responsibilities Split

Main chat / coordinator

• Pick topology based on task type and agent count
• Spawn agents in batched messages (one batch per dependency level)
• Read return values; forward data between levels
• Aggregate consensus scores for Loop 2 gate

Individual agents

• Execute scoped task
• Return: deliverable list, confidence score, brief summary
• Never invoke another agent
• Never wait on external signal

Related Skills

• cfn-loop-orchestration-v2/: full Loop 2/3 orchestration, gate checks, Product Owner decisions
• cfn-agent-lifecycle/: agent lifecycle (spawn, audit, transcript) and SQLite tracking via AGENT_LIFECYCLE_DB. Consolidates the former cfn-agent-spawning/.

History

• 2026-05-13: Created. Consolidates salvaged content from deprecated docs (coordinator-patterns.md, spawn-pattern-examples.md, coordinator-feedback-pattern.md) which described Redis lpush/blpop coordination. Originals archived to docs/archive/.