{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-08-001",
    "source_story_id": "tmp-lead-programmatic-tool-calling",
    "edition_id": "mp-2026-08-08-morning-0030",
    "edition_url": "https://themachinepress.com/edition/2026-08-08",
    "position": 1,
    "story_type": "lead",
    "section": "developer-tools",
    "editorial_classification": "editorial",
    "headline": "Code Became the Better Tool Call",
    "slug": "code-became-the-better-tool-call",
    "dek": "Across fourteen models, typed programs usually matched or beat rigid JSON calls—and held up better when the surrounding context decayed.",
    "summary": "Across fourteen models, typed programs usually matched or beat rigid JSON calls—and held up better when the surrounding context decayed.",
    "body_text": "Researchers compared programmatic tool calling, where a model writes against typed Python stubs and receives execution results in one turn, with native JSON tool calls across 14 language models on Berkeley Function Calling Leaderboard v4. The programmatic approach matched or exceeded JSON in 11 models; the authors report a 10.6 percent gain for the GPT-5.6 family, parity or improvement for 13 models under parallel fan-out, and stable performance in a context-rot test where the JSON baseline fell 2.3 percent on average. These are author-reported benchmark results, not proof that arbitrary generated code is safe or cheaper to execute in production.",
    "why_it_matters": "Across fourteen models, typed programs usually matched or beat rigid JSON calls—and held up better when the surrounding context decayed.",
    "limitations": [
      "These are author-reported benchmark results, not proof that arbitrary generated code is safe or cheaper to execute in production."
    ],
    "importance": 10,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-08-001/code-became-the-better-tool-call",
    "json_url": "https://themachinepress.com/story/mp-2026-08-08-001.json",
    "first_published_at": "2026-08-08T09:00:00.000-04:00",
    "modified_at": "2026-08-08T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-08-001-001",
        "text": "Across fourteen models, typed programs usually matched or beat rigid JSON calls—and held up better when the surrounding context decayed.",
        "source_ids": [
          "source-2026-08-08-001"
        ],
        "qualification": "These are author-reported benchmark results, not proof that arbitrary generated code is safe or cheaper to execute in production."
      }
    ],
    "source_ids": [
      "source-2026-08-08-001"
    ],
    "tags": [
      "AI agents",
      "tool calling",
      "program synthesis",
      "evaluation"
    ],
    "image_url": "https://themachinepress.com/issues/2026-08-08/lead-programmatic-tool-calling.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-08-001",
      "title": "arXiv preprint 2608.06370",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.06370",
      "canonical_url": "https://arxiv.org/abs/2608.06370",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": null,
      "accessed_at": "2026-08-08T08:24:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-08-001-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Code Became the Better Tool Call",
    "publisher": "The Machine Press",
    "published_at": "2026-08-08T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-08-001/code-became-the-better-tool-call"
  }
}