{ "schema_version": 1, "builder_plan": { "scenario": "An agent reads external documents or tool responses that may contain malicious instructions.", "change": "Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.", "baseline": "Your existing checks with the same tasks, permissions, and tool responses.", "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": "Successful attacks missed, legitimate actions blocked, detection delay, and review cost.", "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": "2609.20822", "title": "Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation", "summary": "SafeHarness replaces free-form code generation by a language-model controller with a decomposed pipeline in which the agent plans waypoint routes over grounded obstacle bounding boxes, verifies and replans before executing, and selects obstacle-aware contact positions. The cost is added planning, verification, and replanning rounds per task, increasing latency and API token consumption, plus a perception stack reliable enough to produce accurate bounding boxes. It can fail when bounding-box grounding is inaccurate, when candidate routes are all infeasible, or when dynamic obstacles invalidate the verified plan between planning and contact execution.", "rationale": "The method requires a physical robot manipulation setup that the reader does not operate, but its core lesson—forcing stated safety constraints into an explicit plan-verify-replan harness rather than relying on prompt instructions—is worth tracking for future constraint-critical agent deployments.", "technique": "Obstacle-aware safety harness for coding agents", "family": "seguranca_e_guardrails", "recommendation": "observar", "published": "2026-09-17", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.20822", "independent_implementations": 1, "total_implementations": 2, "stars": 217, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "SafeHarness replaces free-form code generation by a language-model controller with a decomposed pipeline in which the agent plans waypoint routes over grounded obstacle bounding boxes, verifies and replans before executing, and selects obstacle-aware contact positions. The cost is added planning, verification, and replanning rounds per task, increasing latency and API token consumption, plus a perception stack reliable enough to produce accurate bounding boxes. It can fail when bounding-box grounding is inaccurate, when candidate routes are all infeasible, or when dynamic obstacles invalidate the verified plan between planning and contact execution.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.20822", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "SafeHarness attains 71.9% task success and 87.5% collision avoidance, 2.3x and 1.5x the same agent without harnesses, and +6.5%/+27.0% over prior SOTA in a constrained robot manipulation benchmark.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.20822", "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": "" }