CAUM is an innovative structural observation layer designed specifically for AI agents. It enables the detection of loops, stagnation, and inefficient compute usage during autonomous agent execution, all without the need to interpret prompts or payloads.

Key Features

• Behavioral Monitoring: CAUM tracks how an agent operates, evaluating tool call diversity, trajectory geometry, and regime transitions.
• Independent Recording: The core functionality is to record the actions of agents without making decisions on their behalf.

Performance Insights

With validation on 80,036 real agent sessions from the nebius/SWE-agent-trajectories dataset, CAUM demonstrates significant efficacy:

Cross-model validation shows consistent results across various models without the requirement for retraining:

Operational Mechanism

CAUM processes agent activity through:

Agent Steps → SBERT Embeddings → Trajectory Analysis → Regime Classification → UDS Score + Attestation

It leverages five structural signals for comprehensive analysis:

• TCR: Tool Coherence Ratio - Assessing tool diversity.
• ESR: Execution Substance Ratio - Measuring meaningfulness of steps.
• SCI: Structural Coherence Index - Evaluating trajectory progress.
• ZT Similarity: Zero-Trust cosine similarity between consecutive steps.
• Regime: Classifying activity into EXPLORER, GRIND, STAGNATION, or LOOP.

Example Output

{
  "uds": 0.73,           # Unified Dynamic Score [0-1]
  "tier": "T2",          # Tier level, T1 (healthy) → T5 (critical)
  "dominant_regime": "EXPLORER",
  "shields_fired": 0,
  "advisory": {"level": 0, "level_name": "CLEAR"}
}

Utilization Scenarios

Use CAUM to:

1. Analyze step-by-step actions of agents in real-time.
2. Score JSONL files directly for batch analysis.

Quick Examples

To analyze steps:

from caum_monitor_v10 import ZeroTrustAuditor

auditor = ZeroTrustAuditor()

auditor.push("file_search", "find authentication.py in /src")
auditor.push("file_reader", "reading JWT validation at line 42")
auditor.push("code_editor", "fixed algorithm parameter in jwt.decode")
auditor.push("test_runner", "pytest — 12 passed in 1.2s")
auditor.push("submit", "PR #142 submitted")

report = auditor.finalize()
print(f"UDS={report['uds']} tier={report['tier']} regime={report['dominant_regime']}")

REST API Request Example

curl -X POST https://caum-observation-production.up.railway.app/v1/analyze \
  -H "Content-Type: application/json" \
  -d '{"trajectory": [{"tool": "search", "content": "..."}, ...]}'

Behavioral Regimes Overview

CAUM categorizes agent behavior into various regimes:

Deployment Flexibility

Choose appropriate deployment modes:

• Forensic: Batch analysis of past sessions for compliance.
• Live API: Real-time monitoring with alerting.
• Enterprise SDK: Integrated within existing infrastructure with no data transfer.

For further exploration or to test capabilities, visit caum.systems/upload for a free structural observation report.