{ "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": "2609.05736", "title": "Beyond Prompts: Measuring and Optimizing LLM Tool-Agent Harnesses", "summary": "The method replaces ad hoc manual prompt and tool-interface tuning with a budgeted search over prompts and guarded tool-boundary middleware edits around a fixed model, requiring no retraining. Costs are optimizer compute and evaluation runs during search, plus ongoing middleware complexity and per-call logging overhead; the paper itself logs cost diagnostics rather than claiming cost reduction. The main failure mode is selecting brittle harness updates that generalize poorly, which is why the protocol mandates worst-condition lift, repeatability, and the conservative RelLift95 metric alongside mean lift.", "rationale": "Harness optimization around a fixed model matches constrained infrastructure directly, but the reported gains are benchmark-specific and the paper's own reliability caveat means the reader must validate lifts on their workload before production adoption.", "technique": "Fixed-model harness optimization with failure-surface routing", "family": "uso_de_ferramenta", "recommendation": "testar", "published": "2026-09-04", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.05736", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces ad hoc manual prompt and tool-interface tuning with a budgeted search over prompts and guarded tool-boundary middleware edits around a fixed model, requiring no retraining. Costs are optimizer compute and evaluation runs during search, plus ongoing middleware complexity and per-call logging overhead; the paper itself logs cost diagnostics rather than claiming cost reduction. The main failure mode is selecting brittle harness updates that generalize poorly, which is why the protocol mandates worst-condition lift, repeatability, and the conservative RelLift95 metric alongside mean lift.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.05736", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "PRISM reports mean held-out lifts of 14.2, 14.9, and 10.1 percentage points on BFCL multi-round, tau2-Retail, and tau2-Telecom, with positive RelLift95 on all three; the ablation attributes the margin mainly to failure-surface routing and the edit-pattern constraint.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.05736", "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": [ { "anchor": "S3.Ex2", "label": "", "section": "3 Evaluation Protocol and Metrics", "role": "metric", "latex": "\\mathrm{RelLift}_{\\gamma}(B)=Q_{1-\\gamma}(L_{B}),", "mathml": "RelLiftγ​(B)=Q1−γ​(LB),", "context": "asks how much held-out lift a procedure can reliably obtain under a fixed pre-scorecard search budget. For run , let be the pre-scorecard search cost under a declared accounting convention, covering search-time activity (proposal, candidate evaluation, filtering, repair, and selection) before final scorecard evaluation; the final held-out scorecard evaluation is always excluded, and the components a given convention prices are stated with it (ours are defined in Appendix F). Let be the mean cost over runs. Under budget , the number of affordable independent optimizer runs is . Each run returns a harness and a pre-scorecard selection statistic (gate score, validation score, or the optimizer’s internal pre-scorecard criterion) computed without access to . Under budget we select over the affordable runs, and the held-out lift of the selected harness is . Our budgeted reliable-lift metric is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.Ex1", "label": "", "section": "3 Evaluation Protocol and Metrics", "role": "baseline", "latex": "\\Delta_{i,m}=S_{b,m}(h_{i},D_{\\mathrm{score},i})-S_{b,m}(h_{0},D_{\\mathrm{score},i}),", "mathml": "Δi,m=Sb,m​(hi,Dscore,i)−Sb,m​(h0,Dscore,i),", "context": "Let h0 be the baseline harness. For run i, let hi be the harness returned before scorecard evaluation. The held-out scorecard lift is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S4.Ex3", "label": "", "section": "4 Experiments", "role": "proposed_method", "latex": "G_{i}=\\big(S_{\\mathrm{gate}}(h_{i},D_{\\mathrm{gate}}),1-\\mathrm{Stuck}(h_{i},D_{\\mathrm{gate}})\\big).", "mathml": "Gi=(Sgate​(hi,Dgate),1−Stuck⁡(hi,Dgate)).", "context": "For each benchmark–optimizer pair, we observe n≤16 run tuples {(Ci,Gi,Δi)}i=1n and apply the RelLift95​(B) estimator of Section 3 with 5,000 bootstrap resamples; the resamples reduce Monte Carlo error and do not increase the number of independent optimizer runs beyond the n observed. In our implementation, the pre-scorecard statistic Gi, computed identically for every optimizer’s returned harness, orders candidates by gate-split pass rate, breaking ties by reliability:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-26", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }