F037: Collective Innovation Protocol

Status: Proposed Target: v1.0.0 (Multi-Agent Cognition Network) Source: README roadmap, Cisco Outshift IoC Depends on: F036 (Reasoning Continuity), F031 (Source Trust Scoring)

Overview

Enable multiple agents to collaboratively reason about novel problems. Instead of one agent making a decision in isolation, a group of agents contribute perspectives, challenge assumptions, and converge on a solution - with the full deliberation captured as a structured multi-agent trace.

Problem

Current CSTP is single-agent: one agent queries, checks guardrails, and records. Even with sub-agents, each decides independently. There’s no protocol for structured multi-agent deliberation where agents explicitly build on, challenge, or refine each other’s reasoning.

Concept

Deliberation Session

{
  "method": "cstp.openDeliberation",
  "params": {
    "topic": "Should we adopt HSM architecture for long-context processing?",
    "category": "architecture",
    "stakes": "high",
    "participants": ["emerson", "minski", "code-reviewer"],
    "protocol": "structured_debate"
  }
}

Contribution Types

Type	Purpose	Example
propose	Initial proposal	“We should use HSM because…”
support	Add supporting evidence	“MIT research confirms 81% improvement”
challenge	Raise concerns	“HSM hasn’t been tested at scale”
synthesize	Combine perspectives	“HSM for parallel, attention for sequential”
vote	Signal position	Confidence-weighted agreement/disagreement

Deliberation Flow

1. OPEN: Emerson proposes topic
2. PROPOSE: Emerson: "Adopt HSM for long-context" (confidence: 0.80)
3. SUPPORT: Minski: "MIT data supports this" (confidence: 0.85)
4. CHALLENGE: CodeReviewer: "No production benchmarks" (confidence: 0.70)
5. SYNTHESIZE: Emerson: "HSM for parallel + attention fallback" (confidence: 0.82)
6. VOTE: All agents signal confidence
7. CLOSE: Decision recorded with full multi-agent trace

Resolution

{
  "result": {
    "decision": "Adopt HSM with attention fallback for sequential tasks",
    "confidence": 0.82,
    "consensusType": "convergent",
    "participantVotes": {
      "emerson": {"position": "support", "confidence": 0.85},
      "minski": {"position": "support", "confidence": 0.80},
      "code-reviewer": {"position": "conditional_support", "confidence": 0.70}
    },
    "dissent": ["No production benchmarks yet - revisit after pilot"]
  }
}

Protocols

Structured Debate

Round-robin: propose -> support/challenge -> synthesize -> vote. Best for high-stakes architectural decisions.

Advisory Panel

One agent proposes, others advise. Proposer makes final call. Best for decisions with a clear owner.

Consensus

All agents must agree above threshold. Best for shared-impact decisions.

API

cstp.openDeliberation

Start a multi-agent deliberation session.

cstp.contribute

Add a contribution (propose/support/challenge/synthesize/vote).

cstp.closeDeliberation

Finalize and record the collective decision.

cstp.getDeliberation

Retrieve full deliberation history.

Integration

• F031 Source Trust weights each agent’s contribution by their track record
• F032 Error Amplification tracks whether collective decisions outperform solo ones
• F036 Reasoning Continuity enables async deliberation across sessions
• F030 Circuit Breakers can trigger mandatory deliberation for high-stakes decisions

Acceptance Criteria

• cstp.openDeliberation RPC method
• cstp.contribute RPC method (all 5 contribution types)
• cstp.closeDeliberation with consensus calculation
• cstp.getDeliberation RPC method
• Multi-agent decision trace stored in CSTP
• MCP tools exposed
• Dashboard: Deliberation timeline view
• At least 3 resolution protocols (debate, advisory, consensus)