Show & Tell: Designing the MeshResultAggregator — Consensus, Confidence, Weighted, and First-Done Aggregation

[web3guru888]

When AgentMesh fans out a task to multiple specialist agents, something has to collect the pieces. That something is MeshResultAggregator.

Here is what it does and how it was designed.

The Problem

Suppose three agents analyze a perception event in parallel:

• ReasoningAgent returns a logical conclusion with confidence 0.91
• KnowledgeAgent returns a factual lookup with confidence 0.87
• SafetyAgent returns a risk score with confidence 0.95

Who wins? How do we detect disagreement? And how does the result reach the Cognitive Blackboard without a race condition?

Four Aggregation Strategies

Strategy	When to use	How it works
CONSENSUS	Discrete outputs (labels, decisions)	Majority vote; ties broken by average confidence
CONFIDENCE	Any output type	Highest-confidence agent wins
WEIGHTED	Numeric outputs	Confidence-weighted average
FIRST_DONE	Speculative execution	Return first success, cancel waiting

The default is CONFIDENCE — it works for any output type and degrades gracefully.

dissent_ratio

Every AggregatedResult carries a dissent_ratio in [0.0, 1.0]:

• 0.0 = all agents agreed (unanimous)
• 1.0 = maximum disagreement

High dissent is a signal worth writing to the Blackboard. If dissent_ratio > 0.4 on a safety-critical task, the CognitiveCycle should pause and request human review.

State Machine

dispatch_task()
      │
      ▼
[COLLECTING]  ←── receive_result() × N
      │
      ├─ quorum reached? ──► [AGGREGATING] ──► _aggregate() ──► bb.post()
      │
      └─ timeout? ──────────► check quorum ──► QuorumNotReached if insufficient

Blackboard Entry

key   = f"mesh.result.{task_id}"
value = AggregatedResult(
    output        = "SafetyAgent conclusion",
    confidence    = 0.95,
    agent_count   = 3,
    dissent_ratio = 0.12,
    latency_ms    = 412.0,
    ...
)
ttl   = 30s   # long enough for downstream consumers
tags  = ["multi_agent", "confidence"]

Subscribers to mesh.result.* (CognitiveCycle, PLN, Safety) receive it instantly via EventBus.

Lock Design

receive_result is on the hot path — called from many concurrent coroutines. The lock is held only for list append and length check. Quorum detection and aggregation happen outside the lock. This keeps contention near zero even with 50+ concurrent agents.

Open Questions

1. Should dissent_ratio trigger an automatic escalation to the Safety module, or just be a passive metric on the Blackboard?
2. For WEIGHTED strategy, should AgentDiscovery reliability scores (Add AgentDiscovery: service registry and health tracking for AgentMesh #150) be factored in alongside confidence, or is confidence alone sufficient?
3. What is the right default timeout? 5s feels long for a 120ms CognitiveCycle tick.

Implementation tracked in #159.