{
  "alt": "Grouped bars for three situations and three models. Nearly all models choose retry after a failed tool. Most avoid finishing when an input is missing, but few ask the user.",
  "caption": "The right answer here comes from how the fault was built. Holding back from finishing is the low bar; asking the user is what would actually unblock the work, though at the first steps an agent may reasonably look around first.",
  "data": [
    {
      "model": "Jev 1.14.0",
      "share": 97.4,
      "situation": "A tool failed and said to try again: chose retry"
    },
    {
      "model": "Jev 1.14.0",
      "share": 80.6,
      "situation": "A required input is still missing: did not finish"
    },
    {
      "model": "Jev 1.14.0",
      "share": 19.4,
      "situation": "A required input is still missing: asked the user"
    },
    {
      "model": "OpenAI Luna",
      "share": 92.1,
      "situation": "A tool failed and said to try again: chose retry"
    },
    {
      "model": "OpenAI Luna",
      "share": 77.4,
      "situation": "A required input is still missing: did not finish"
    },
    {
      "model": "OpenAI Luna",
      "share": 12.9,
      "situation": "A required input is still missing: asked the user"
    },
    {
      "model": "Claude Haiku adapter",
      "share": 92.1,
      "situation": "A tool failed and said to try again: chose retry"
    },
    {
      "model": "Claude Haiku adapter",
      "share": 90.3,
      "situation": "A required input is still missing: did not finish"
    },
    {
      "model": "Claude Haiku adapter",
      "share": 14.5,
      "situation": "A required input is still missing: asked the user"
    }
  ],
  "id": "handling-faults",
  "source": "Constructed agent tasks written for this booklet; one agent (GPT-6 Luna), 200 runs with injected faults, October 2026; unweighted",
  "spec": {
    "category": "situation",
    "colors": {
      "Claude Haiku adapter": "cat3",
      "Jev 1.14.0": "cat1",
      "OpenAI Luna": "cat2"
    },
    "format": {
      "digits": 0
    },
    "group": "model",
    "orientation": "h",
    "sort": "data",
    "value": "share",
    "valueLabel": "Share of checkpoints (%)"
  },
  "subtitle": "Share of checkpoints where each model's next action fits the situation, percent",
  "title": "The models handle a failed tool well and rarely ask for a missing input",
  "type": "bar"
}