{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "An assistant chooses the wrong API or sends arguments your product cannot use.",
    "change": "Apply the proposed tool-selection or argument-generation change in a sandbox.",
    "baseline": "Your current agent with the same tool schemas and task budget.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Correct completed actions, invalid calls, retries, latency, and cost per completed task.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2610.28601",
  "title": "Evolving Agent Interfaces",
  "summary": "The method replaces hand-designed tool interfaces and hand-written usage knowledge for computer-use agents with a repository of interfaces that a meta-agent constructs from unlabeled task trajectories and selects via population-based evolution on held-out validation tasks. The cost is the offline evolution pipeline itself: many synthetic training tasks, full agent rollouts, and a meta-agent pass per iteration, which is significant API or GPU spend even though inference with the resulting repository adds little overhead. What can fail is generalization: interfaces evolved on synthetic tasks may not transfer to the reader's real applications, the open-source repository may be tuned to OSWorld-style benchmarks, and running custom evolution on a single 24 GB GPU with a small budget is likely impractical.",
  "rationale": "The reader can cheaply test the released evolved repository as a drop-in interface layer for computer-use agents, but running the full evolution pipeline on their own domain requires compute and task-generation effort that likely exceeds their constrained infrastructure.",
  "technique": "Evolution of agent tool interfaces from trajectories",
  "family": "uso_de_ferramenta",
  "recommendation": "testar",
  "published": "2026-10-20",
  "as_of": "2026-10-21",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.28601",
  "independent_implementations": 3,
  "total_implementations": 4,
  "stars": 12802,
  "citations": null,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "The method replaces hand-designed tool interfaces and hand-written usage knowledge for computer-use agents with a repository of interfaces that a meta-agent constructs from unlabeled task trajectories and selects via population-based evolution on held-out validation tasks. The cost is the offline evolution pipeline itself: many synthetic training tasks, full agent rollouts, and a meta-agent pass per iteration, which is significant API or GPU spend even though inference with the resulting repository adds little overhead. What can fail is generalization: interfaces evolved on synthetic tasks may not transfer to the reader's real applications, the open-source repository may be tuned to OSWorld-style benchmarks, and running custom evolution on a single 24 GB GPU with a small budget is likely impractical.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.28601",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "On OSWorld V2, the evolved interface repository raises Qwen3.8-27B binary score to 21.3%, 3.8x the same model with fixed interfaces; on ScienceBoard success rises from 41.4% to 58.6%.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.28601",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}