Team Coherence

Keep your AI teams aligned and prevent drift

MeetLoyd's Team Coherence system monitors your AI agent teams to ensure they stay aligned with their intended objectives. Detect and prevent common failure modes like agents undoing each other's work or getting stuck in correction loops.

Overview

The Team Coherence system allows you to:

• Define team objectives and constraints
• Monitor coherence scores in real-time
• Detect drift patterns before they cause problems
• Set budgets to prevent runaway behavior
• Pause teams automatically when issues are detected

Key Concept

Agent failure isn't about intelligence—it's about drift from original intent. Coherence monitoring catches these issues early.

Accessing Team Coherence

Navigate to Governance → Team Coherence in the dashboard, or view the coherence widget on any team's detail page.

Understanding Coherence Score

Each team has a coherence score from 0-100:

Score	Status	Meaning
80-100	Healthy	Team is well-aligned and productive
60-79	Warning	Some drift detected, monitor closely
Below 60	Critical	Significant drift, intervention recommended

Drift Patterns Detected

Oscillation

Agents repeatedly undoing each other's work:

• Agent A makes a change
• Agent B reverts it
• Agent A re-does it

Symptom: High oscillation score, low productivity

Feedback Loops

Circular correction chains:

• Agent A corrects Agent B
• Agent B corrects Agent C
• Agent C corrects Agent A

Symptom: High correction rate, no forward progress

Correction Spikes

Sudden increase in agents correcting each other's work.

Symptom: Elevated corrections per hour

Setting Up Intent State

Define what your team should accomplish:

1. Primary Objective

The main goal for the team:

"Process customer support tickets and resolve issues within SLA"

2. Secondary Objectives

Supporting goals:

• "Maintain customer satisfaction above 4.5/5"
• "Escalate complex issues to human agents"

3. Constraints

Boundaries the team must respect:

Type	Description	Example
Hard	Never violate	"Never share customer data externally"
Soft	Minimize violations	"Prefer email over phone contact"

Coherence Budget

Set limits to prevent runaway behavior:

Limit	Default	Description
Max Concurrent Agents	5	Agents active simultaneously
Max Actions/Minute	30	Rate limit on operations
Max Corrections/Hour	10	Correction threshold
Max Feedback Loop Depth	3	Circular correction limit

Enforcement Modes

Mode	Behavior
Warn	Alert but continue operations
Pause	Auto-pause team for cooldown period
Block	Require human intervention to resume

Drift Events

When issues are detected, drift events are created:

Event Type	Severity	Description
oscillation_detected	Warning/Critical	Agents undoing each other
feedback_loop_detected	Warning/Critical	Circular corrections
correction_spike	Warning/Critical	High correction rate
budget_exceeded	Critical	Rate limits hit

Resolving Events

1. Acknowledge - You've seen it, investigating
2. Resolve - Issue fixed, document what was done
3. Dismiss - False positive, adjust thresholds

Dashboard Views

Tenant Summary

Overview of all monitored teams:

• Teams at risk count
• Average coherence score
• Recent unresolved events

Team Detail

Deep dive into specific team:

• Real-time coherence gauge
• Trend graph (improving/stable/degrading)
• Budget usage meters
• Recent events list

API Integration

Get Team Coherence Summary

GET /api/coherence/teams/:teamId/summary

Response:

{
  "team": {
    "id": "team_xxx",
    "coherenceScore": 85,
    "trend": "stable",
    "isPaused": false
  },
  "metrics": {
    "oscillationScore": 0.1,
    "correctionRate": 2.5,
    "intentAlignment": 92
  },
  "recentEvents": []
}

Pause a Team

POST /api/coherence/teams/:teamId/pause
Content-Type: application/json

{
  "reason": "Investigating high correction rate",
  "durationMinutes": 30
}

Resume a Team

POST /api/coherence/teams/:teamId/resume
Content-Type: application/json

{
  "notes": "Root cause identified and fixed"
}

Get Drift Events

GET /api/coherence/teams/:teamId/drift/events?severity=warning&limit=20

Best Practices

Initial Setup

1. Start with default budgets and thresholds
2. Let the team run for a day to establish baseline
3. Tune thresholds based on observed behavior

Ongoing Monitoring

1. Check coherence scores daily
2. Investigate warnings promptly—don't let drift accumulate
3. Review feedback loops—often indicates unclear responsibilities

When Drift is Detected

1. Pause the team if critical
2. Review recent actions to understand the pattern
3. Clarify objectives or constraints if needed
4. Resume with monitoring

Related Features

• Compliance Verification - Policy verification
• Governance Framework - Overall governance
• Kill Switch - Emergency controls