{
  "id": "agent-reliability-faq",
  "type": "faq",
  "name": "Agent reliability FAQ",
  "summary": "The questions teams actually ask before shipping an autonomous agent: how to stop hallucinated actions against critical APIs, what a benchmark should measure, whether one successful demo means anything, and where to start when there is no eval at all. Answers are bound to sources, not opinion.",
  "locale": "en",
  "tags": [
    "faq",
    "reliability",
    "benchmarks",
    "hallucination"
  ],
  "relations": [
    {
      "rel": "related",
      "target": "measuring-agent-reliability"
    },
    {
      "rel": "related",
      "target": "sandboxed-execution"
    },
    {
      "rel": "related",
      "target": "regression-gating-model-upgrades"
    },
    {
      "rel": "related",
      "target": "choosing-k-for-pass-hat-k"
    }
  ],
  "questions": [
    "How do I stop my agent hallucinating against a critical API?",
    "What should an agent benchmark measure?",
    "My agent passed the demo — is it production-ready?",
    "We have no evals at all. Where do we start?",
    "Which existing agent benchmarks should I look at first?",
    "How do I upgrade the model behind my agent without breaking it?",
    "Is a small eval difference between two agents real?",
    "How many times should I run each task before shipping?"
  ],
  "claims": [
    {
      "id": "c1",
      "text": "Hallucinated actions are caught by grading state, not prose: comparing the final state of the acted-on system against an annotated goal state detects wrong actions that a plausible transcript hides.",
      "sources": [
        "tau-bench"
      ],
      "confidence": 0.9
    },
    {
      "id": "c2",
      "text": "Until verified, an agent belongs in a contained environment; execution-graded evaluation presumes an execution environment the agent cannot escape or corrupt.",
      "sources": [
        "swe-bench"
      ],
      "confidence": 0.85
    },
    {
      "id": "c3",
      "text": "An agent benchmark should measure consistency across repeated trials, not just single-attempt success: pass^k exposes agents that pass once but fail the user who retries.",
      "sources": [
        "tau-bench"
      ],
      "confidence": 0.9
    },
    {
      "id": "c4",
      "text": "Capability trend and reliability are different measurements: the length of tasks agents can complete has grown exponentially, which says nothing about consistency on any one task.",
      "sources": [
        "metr"
      ],
      "confidence": 0.85
    },
    {
      "id": "c5",
      "text": "Before building a benchmark, survey the shelf: a public catalog tracks 69 agentic benchmarks and the field's move from measuring capability to measuring agency.",
      "sources": [
        "santismm-eureka"
      ],
      "confidence": 0.85
    },
    {
      "id": "c6",
      "text": "Model swaps run on the provider's calendar: deprecated models get as little as 60 days between notice and retirement, after which requests fail — so a frozen regression suite for model upgrades is standing infrastructure, not a one-off project.",
      "sources": [
        "anthropic-model-deprecations"
      ],
      "confidence": 0.9
    },
    {
      "id": "c7",
      "text": "A score difference smaller than the eval's noise floor is not a finding: eval questions behave as samples from a larger population and comparisons need standard errors — reported noise from formatting choices alone has reached about 5 points.",
      "sources": [
        "error-bars-evals",
        "anthropic-eval-challenges"
      ],
      "confidence": 0.85
    },
    {
      "id": "c8",
      "text": "Run each task as many times as it runs unreviewed in production, and read the right metric: all-of-k reliability collapses with repetition (pass^8 under 25% where single attempts clear 50%), while any-of-k capability climbs with it (28.8% to 70.2% from 1 to 100 samples) — so the same suite supports opposite conclusions depending on which one you quote.",
      "sources": [
        "tau-bench",
        "codex-humaneval"
      ],
      "confidence": 0.9
    }
  ],
  "takeaways": [],
  "faqs": [
    {
      "q": "How do I stop my agent hallucinating against a critical API?",
      "claim": "c1"
    },
    {
      "q": "Where should an unproven agent run?",
      "claim": "c2"
    },
    {
      "q": "What should an agent benchmark measure?",
      "claim": "c3"
    },
    {
      "q": "My agent passed the demo — is it production-ready?",
      "claim": "c4"
    },
    {
      "q": "Which existing agent benchmarks should I look at first?",
      "claim": "c5"
    },
    {
      "q": "How do I upgrade the model behind my agent without breaking it?",
      "claim": "c6"
    },
    {
      "q": "Is a small eval difference between two agents real?",
      "claim": "c7"
    },
    {
      "q": "How many times should I run each task before shipping?",
      "claim": "c8"
    }
  ],
  "evidence_tier": "secondary",
  "evidence": {
    "level": "industry_observation",
    "source_types": [
      "paper",
      "industry_observation"
    ]
  },
  "moat_flag": false,
  "winning_edge": "Answers are the questions buyers ask, phrased as they ask them, each bound to a source with a number in it — not a marketing FAQ and not a survey paper.",
  "confidence": 0.85,
  "last_verified": "2026-08-08",
  "canonical_url": "https://agentreliability.dev/k/agent-reliability-faq",
  "api_url": "https://agentreliability.dev/api/k/agent-reliability-faq.json",
  "jsonld": {
    "@context": "https://schema.org",
    "name": "Agent reliability FAQ",
    "description": "The questions teams actually ask before shipping an autonomous agent: how to stop hallucinated actions against critical APIs, what a benchmark should measure, whether one successful demo means anything, and where to start when there is no eval at all. Answers are bound to sources, not opinion.",
    "url": "https://agentreliability.dev/k/agent-reliability-faq",
    "license": "https://spdx.org/licenses/CC-BY-4.0.html",
    "dateModified": "2026-08-08",
    "citation": [
      {
        "@type": "CreativeWork",
        "name": "EUREKA — applied AI, agents and robotics knowledge platform",
        "url": "https://santismm.com"
      },
      {
        "@type": "CreativeWork",
        "name": "Challenges in evaluating AI systems",
        "url": "https://www.anthropic.com/news/evaluating-ai-systems"
      },
      {
        "@type": "CreativeWork",
        "name": "Anthropic model deprecations (Claude API documentation)",
        "url": "https://platform.claude.com/docs/en/docs/about-claude/model-deprecations"
      },
      {
        "@type": "CreativeWork",
        "name": "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?",
        "url": "https://arxiv.org/abs/2310.06770"
      },
      {
        "@type": "CreativeWork",
        "name": "tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains",
        "url": "https://arxiv.org/abs/2406.12045"
      },
      {
        "@type": "CreativeWork",
        "name": "Evaluating Large Language Models Trained on Code",
        "url": "https://arxiv.org/abs/2107.03374"
      },
      {
        "@type": "CreativeWork",
        "name": "METR — Model Evaluation & Threat Research",
        "url": "https://metr.org/"
      },
      {
        "@type": "CreativeWork",
        "name": "Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations",
        "url": "https://arxiv.org/abs/2411.00640"
      }
    ],
    "author": {
      "@type": "Person",
      "name": "Santiago Santa María Morales",
      "jobTitle": "practitioner — harness engineering and agent evaluation in production"
    },
    "@type": "FAQPage",
    "mainEntity": [
      {
        "@type": "Question",
        "name": "How do I stop my agent hallucinating against a critical API?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Hallucinated actions are caught by grading state, not prose: comparing the final state of the acted-on system against an annotated goal state detects wrong actions that a plausible transcript hides."
        }
      },
      {
        "@type": "Question",
        "name": "Where should an unproven agent run?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Until verified, an agent belongs in a contained environment; execution-graded evaluation presumes an execution environment the agent cannot escape or corrupt."
        }
      },
      {
        "@type": "Question",
        "name": "What should an agent benchmark measure?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "An agent benchmark should measure consistency across repeated trials, not just single-attempt success: pass^k exposes agents that pass once but fail the user who retries."
        }
      },
      {
        "@type": "Question",
        "name": "My agent passed the demo — is it production-ready?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Capability trend and reliability are different measurements: the length of tasks agents can complete has grown exponentially, which says nothing about consistency on any one task."
        }
      },
      {
        "@type": "Question",
        "name": "Which existing agent benchmarks should I look at first?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Before building a benchmark, survey the shelf: a public catalog tracks 69 agentic benchmarks and the field's move from measuring capability to measuring agency."
        }
      },
      {
        "@type": "Question",
        "name": "How do I upgrade the model behind my agent without breaking it?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Model swaps run on the provider's calendar: deprecated models get as little as 60 days between notice and retirement, after which requests fail — so a frozen regression suite for model upgrades is standing infrastructure, not a one-off project."
        }
      },
      {
        "@type": "Question",
        "name": "Is a small eval difference between two agents real?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "A score difference smaller than the eval's noise floor is not a finding: eval questions behave as samples from a larger population and comparisons need standard errors — reported noise from formatting choices alone has reached about 5 points."
        }
      },
      {
        "@type": "Question",
        "name": "How many times should I run each task before shipping?",
        "acceptedAnswer": {
          "@type": "Answer",
          "text": "Run each task as many times as it runs unreviewed in production, and read the right metric: all-of-k reliability collapses with repetition (pass^8 under 25% where single attempts clear 50%), while any-of-k capability climbs with it (28.8% to 70.2% from 1 to 100 samples) — so the same suite supports opposite conclusions depending on which one you quote."
        }
      }
    ]
  }
}