Skip to main content

​
Overview

While pairwise value coherence checks compatibility between two agents, fleet coherence extends this to N agents simultaneously. It computes all C(n,2) pairwise scores and derives fleet-level insights:
  • Fleet Score — Mean of all pairwise coherence scores
  • Outlier Detection — Agents significantly below the fleet mean
  • Cluster Analysis — Groups of compatible agents
  • Divergence Report — Values where agents disagree

​
Fleet Score

The fleet score is the mean of all pairwise coherence scores: 
fleet_score = mean(score(A,B) for all pairs A,B in fleet)
Additionally, the weakest link (minimum pairwise score) identifies the most strained relationship in the fleet.
Fleet ScoreInterpretation
≥ 0.85Strong fleet alignment — safe for autonomous coordination
0.70–0.85Good alignment with minor divergences
0.50–0.70Moderate alignment — review outliers before coordination
< 0.50Poor alignment — significant value conflicts present

​
Outlier Detection

Per-agent mean pairwise scores are computed. An agent is flagged as an outlier if its mean score is more than 1 standard deviation below the fleet mean: 
agent_mean = mean(score(agent, peer) for all peers)
fleet_mean = mean(all agent_means)
stddev = std(all agent_means)

is_outlier = agent_mean < fleet_mean - 1.0 * stddev
Outlier agents are reported with their primary conflict values — the specific value disagreements causing friction.

​
Cluster Analysis

Agents are grouped into clusters using connected components on a compatibility graph:
  1. Build adjacency graph where an edge exists if compatible === true (score ≥ 0.70, no conflicts)
  2. Find connected components via BFS
  3. For each cluster, compute internal coherence and shared values
Clusters reveal natural groupings — for example, in an incident response fleet, monitoring and classification agents may form one cluster while remediation agents form another.

​
Divergence Report

For each value in the union of all declared values, the report shows:
  • Which agents declare it
  • Which agents are missing it
  • Which agents have it in conflicts_with
  • Estimated impact on fleet score
Values are sorted by impact, helping fleet operators prioritize alignment work.

​
SDK Usage

​
TypeScript

import { checkFleetCoherence } from '@mnemom/agent-alignment-protocol';
import type { FleetCoherenceResult } from '@mnemom/agent-alignment-protocol';

const result: FleetCoherenceResult = checkFleetCoherence([
  { agentId: 'sentinel', card: sentinelCard },
  { agentId: 'triage', card: triageCard },
  { agentId: 'patch', card: patchCard },
  { agentId: 'herald', card: heraldCard },
]);

console.log(`Fleet score: ${(result.fleet_score * 100).toFixed(1)}%`);
console.log(`Weakest link: ${(result.min_pair_score * 100).toFixed(1)}%`);
console.log(`Outliers: ${result.outliers.map(o => o.agent_id).join(', ') || 'none'}`);
console.log(`Clusters: ${result.clusters.length}`);

​
Python

from aap import check_fleet_coherence

result = check_fleet_coherence([
    {"agent_id": "sentinel", "card": sentinel_card},
    {"agent_id": "triage", "card": triage_card},
    {"agent_id": "patch", "card": patch_card},
    {"agent_id": "herald", "card": herald_card},
])

print(f"Fleet score: {result.fleet_score:.1%}")
print(f"Outliers: {[o.agent_id for o in result.outliers]}")
print(f"Clusters: {len(result.clusters)}")

​
Enterprise API

For enterprise organizations, fleet coherence is available via the REST API: 
GET /v1/orgs/{org_id}/coherence
Authorization: Bearer <token>
Results are cached for 5 minutes. Requires the nway_coherence feature flag (Enterprise plan).

​
Use Cases

  • Fleet Management — Monitor overall value alignment across all agents in an organization
  • Compliance — Identify agents whose values diverge from organizational standards
  • Incident Response — Verify response team coherence before coordinating autonomous actions
  • Onboarding — Assess how a new agent’s alignment card fits with the existing fleet

​
See Also