AI Agent Performance Auditor

Measure whether an AI agent is reliable, economical, and safe to promote to production. Supply observed runs from one release or model and optionally a baseline cohort. The Actor returns a structured health report for engineering, product, FinOps, and AI-operations teams.

What you get

• Success and failure rates
• Cost per run and cost per successful run
• Revenue, gross profit, and gross margin when revenue data is supplied
• Average and p95 latency
• Average output-quality score
• Model and version cohort comparisons
• Regression detection against a baseline
• Prioritized operational recommendations

Who this is for

• AI engineering teams evaluating a release before rollout
• Platform teams comparing models or agent versions
• FinOps teams monitoring cost per successful outcome
• AI product owners tracking reliability, quality, latency, and margin
• Automated CI or observability workflows that need predictable JSON

Example input

{
  "runs": [{
    "id": "run-101",
    "success": true,
    "costUsd": 0.04,
    "revenueUsd": 0.20,
    "latencyMs": 1200,
    "qualityScore": 92,
    "version": "1.1",
    "model": "model-a"
  }],
  "baselineRuns": [{
    "id": "run-001",
    "success": true,
    "costUsd": 0.06,
    "revenueUsd": 0.20,
    "latencyMs": 1700,
    "qualityScore": 84,
    "version": "1.0",
    "model": "model-a"
  }],
  "thresholds": {
    "successRateDropPoints": 5,
    "qualityDropPoints": 5,
    "latencyIncreasePercent": 20,
    "costPerSuccessIncreasePercent": 20,
    "marginDropPoints": 5
  }
}

Each run accepts success or status. Optional fields are costUsd, revenueUsd, latencyMs, qualityScore from 0–100, version, model, and timestamp.

Output

One report is written to the default dataset. It includes a CI-friendly releaseDecision (GO, REVIEW, or HOLD), healthScore, status, aggregate metrics, regressions, byVersion, byModel, and recommendations. Export it as JSON, CSV, or Excel, or consume it through the Apify API.

Pricing

The Actor charges one agent-performance-audit event per successfully generated report. Invalid input fails before a paid report is produced. Platform usage is included; use Apify's maximum-charge control for an additional run guardrail.

Responsible use

Only submit operational data you are authorized to process. Do not include secrets, personal data, private prompts, or raw customer content. Metrics describe the supplied observations and do not guarantee future revenue or model quality.