{
  "format": "minglabs/v1",
  "surface": "insights-article",
  "slug": "agents-grading-themselves",
  "kind": "article",
  "subtype": "position-paper",
  "url": "https://www.minglabs.com/insights/articles/agents-grading-themselves",
  "htmlUrl": "https://www.minglabs.com/insights/articles/agents-grading-themselves",
  "partOf": "https://www.minglabs.com/insights/articles",
  "title": "Your agent thinks it's doing great work. It isn't.",
  "subtitle": "The most dangerous failure in a hybrid organisation isn't bad output. It's confident output the agent grades well itself.",
  "hook": "Agents grade themselves generously. Left alone, an agent marks its work shipped, writes itself a clean status update, and moves on — while the actual work degrades quietly. The gap between an agent's confidence and the real quality of its output is the most dangerous failure mode in a hybrid organisation, because it passes the only check most teams run: the agent's own.",
  "summary": "We built a research agent that rated its own output as solid while the client called it shallow. The diagnosis: the most common request was owned by no specific skill, so the agent free-formed it and graded against its own loose standard. The fix is structural — make the standard mandatory and machine-checked rather than asking the agent to 'be thorough,' keep a human gate until the system enforces it, and let agents earn autonomy step by step.",
  "category": "Hybrid Organisation",
  "pillar": "Hybrid Organisation",
  "author": {
    "name": "Sebastian Mueller",
    "role": "Founding Partner, MING Labs"
  },
  "datePublished": "2026-05-27",
  "dateModified": "2026-05-27",
  "freshness": {
    "updated": "May 2026",
    "nextReview": "November 2026"
  },
  "evidenceTier": "proprietary",
  "confidence": "B",
  "sources": [
    {
      "id": "S1",
      "title": "Internal MING Labs agent rebuild and validation log (enterprise research agent)",
      "publisher": "MING Labs",
      "date": "2026-05-27",
      "supports": [
        "Shallow-output incident",
        "Scoping root cause",
        "Structural fix",
        "Retained human gate"
      ]
    },
    {
      "id": "S2",
      "title": "MING Labs operating principles — \"evidence over assertion: execute, don't narrate\"",
      "publisher": "MING Labs",
      "date": "2026-04-01",
      "supports": [
        "Principle definition",
        "Mandatory self-review"
      ]
    },
    {
      "id": "S3",
      "title": "MING Labs autonomy-level framework and learning protocol",
      "publisher": "MING Labs",
      "date": "2026-02-15",
      "supports": [
        "Earn-autonomy-like-a-hire model",
        "'Correct the same thing twice' learning metric"
      ]
    }
  ],
  "faqs": [
    {
      "q": "Why can't you trust an AI agent's own quality assessment?",
      "a": "Agents grade themselves against whatever standard is baked into their defaults, and they have no internal signal when that standard is too low. An agent can produce shallow work and rate it as good — honestly — which means its confidence tells you nothing about actual quality."
    },
    {
      "q": "Is shallow agent output a model problem?",
      "a": "Usually not. In our case the agent was fully capable of deep, sourced work; the issue was that the most common task had no dedicated skill, so depth wasn't the default. The fix was structural — making depth and traceability the default — not a more capable model or a better prompt."
    },
    {
      "q": "How do you stop agents from shipping low-quality work?",
      "a": "Treat verification as a job, not a setting. Make standards mandatory and machine-checked rather than asking the agent to 'be thorough,' keep a human gate until the check is enforced by the system, and let agents earn autonomy step by step like a new hire. The signal it's working: you never correct the same thing twice."
    }
  ],
  "relatedConceptSlugs": [
    "what-is-hybrid-organisation",
    "what-is-the-abc-framework"
  ],
  "relatedArticleSlugs": [
    "we-fired-an-ai-agent",
    "hire-dont-deploy"
  ]
}