{
  "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": "2610.28201",
  "title": "MaD-RSI: Model-as-Data for Recursive Self-Improvement of Agent Harnesses",
  "summary": "MaD-RSI replaces single-model recursive self-improvement of agent harnesses, in which a harness is overfit to one LLM and must be re-optimized per model, with joint optimization of one shared harness using failure trajectories from several heterogeneous LLMs. The cost is a more complex training pipeline that must run and collect trajectories from multiple models during optimization, plus API or inference spend for each participating model, which is non-trivial even if lower than per-model re-optimization. It can fail if the shared harness converges to an average that underperforms a model-specific harness on the reader's actual production model, if the reader only uses one or two models (removing the diversity signal the method depends on), or if trajectory collection across API providers is rate-limited or expensive.",
  "rationale": "The method targets harness portability across models, which matters mainly when switching or mixing LLMs at scale, and its unquantified cost and quality claims plus a multi-model training loop make it premature for a constrained-infrastructure team to adopt before independent validation.",
  "technique": "Multi-model shared agent harness optimization",
  "family": "uso_de_ferramenta",
  "recommendation": "observar",
  "published": "2026-10-20",
  "as_of": "2026-10-21",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.28201",
  "independent_implementations": 2,
  "total_implementations": 3,
  "stars": 132,
  "citations": null,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "MaD-RSI replaces single-model recursive self-improvement of agent harnesses, in which a harness is overfit to one LLM and must be re-optimized per model, with joint optimization of one shared harness using failure trajectories from several heterogeneous LLMs. The cost is a more complex training pipeline that must run and collect trajectories from multiple models during optimization, plus API or inference spend for each participating model, which is non-trivial even if lower than per-model re-optimization. It can fail if the shared harness converges to an average that underperforms a model-specific harness on the reader's actual production model, if the reader only uses one or two models (removing the diversity signal the method depends on), or if trajectory collection across API providers is rate-limited or expensive.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.28201",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "The abstract claims improved generalization over existing RSI methods and reduced training costs, but reports no specific numbers; gains would need verification in the full paper's experimental section.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.28201",
      "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": ""
}