{
  "title": "Harness engineering research source ledger",
  "review_date": "2026-09-10",
  "scope": "Primary research and implementation sources supporting the four-post September series; selective coverage, not an exhaustive literature census.",
  "method": "Primary papers and project materials reviewed; published benchmark JSON independently recalculated. No original agent trials rerun. Future experiments are proposals.",
  "sources": [
    {
      "id": "long-running",
      "title": "Effective harnesses for long-running agents",
      "author_or_publisher": "Justin Young / Anthropic",
      "publication_or_revision_date": "2025-11-26",
      "url": "https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents",
      "evidence_class": "first-party engineering report",
      "scope_and_caveat": "Incremental work and durable continuation artifacts; workload-specific experience.",
      "accessed": "2026-09-10"
    },
    {
      "id": "infrastructure",
      "title": "Quantifying infrastructure noise in agentic coding evals",
      "author_or_publisher": "Gian Segato / Anthropic",
      "publication_or_revision_date": "2026-02-05",
      "url": "https://www.anthropic.com/engineering/infrastructure-noise",
      "evidence_class": "first-party experiment",
      "scope_and_caveat": "Resource enforcement is an experimental variable.",
      "accessed": "2026-09-10"
    },
    {
      "id": "deepagents",
      "title": "Improving Deep Agents with harness engineering",
      "author_or_publisher": "Vivek Trivedy / LangChain",
      "publication_or_revision_date": "2026-02-17",
      "url": "https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering",
      "evidence_class": "first-party engineering report",
      "scope_and_caveat": "Fixed-model trace-guided harness intervention; not independently rerun.",
      "accessed": "2026-09-10"
    },
    {
      "id": "application-harness",
      "title": "Harness design for long-running application development",
      "author_or_publisher": "Prithvi Rajasekaran / Anthropic",
      "publication_or_revision_date": "2026-03-24",
      "url": "https://www.anthropic.com/engineering/harness-design-long-running-apps",
      "evidence_class": "first-party engineering report",
      "scope_and_caveat": "Planning, generation, evaluation and long-running applications.",
      "accessed": "2026-09-10"
    },
    {
      "id": "meta-harness",
      "title": "Meta-Harness: End-to-End Optimization of Model Harnesses",
      "author_or_publisher": "Yoonho Lee and colleagues",
      "publication_or_revision_date": "2026-03-30",
      "url": "https://arxiv.org/abs/2603.28052",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Executable harness search with previous candidate evidence.",
      "accessed": "2026-09-10"
    },
    {
      "id": "managed-agents",
      "title": "Scaling Managed Agents: Decoupling the brain from the hands",
      "author_or_publisher": "Anthropic",
      "publication_or_revision_date": "2026-04-08",
      "url": "https://www.anthropic.com/engineering/managed-agents",
      "evidence_class": "first-party architecture report",
      "scope_and_caveat": "Separate session, harness, sandbox; remove stale model workarounds.",
      "accessed": "2026-09-10"
    },
    {
      "id": "update-benefit",
      "title": "Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents",
      "author_or_publisher": "Minhua Lin and colleagues",
      "publication_or_revision_date": "2026-05-28",
      "url": "https://arxiv.org/html/2605.30621v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Updating and benefiting from updates are separate capabilities.",
      "accessed": "2026-09-10"
    },
    {
      "id": "harnesscompass",
      "title": "HarnessCompass: Guiding Automatic Harness Evolution toward Generalizable and Effective Agent Harnesses",
      "author_or_publisher": "Luan Zhang and colleagues",
      "publication_or_revision_date": "2026-08-03",
      "url": "https://arxiv.org/html/2608.01918v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Component-wise constrained evolution and transfer claims.",
      "accessed": "2026-09-10"
    },
    {
      "id": "harnessopt",
      "title": "HarnessOpt-Bench: Evaluating LLMs at Harness Optimization",
      "author_or_publisher": "Varun Ursekar and colleagues / Scale AI",
      "publication_or_revision_date": "2026-08-06",
      "url": "https://arxiv.org/html/2608.06301v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Metered search with protected test scoring; 111 scored runs.",
      "accessed": "2026-09-10"
    },
    {
      "id": "harnessrisk",
      "title": "HarnessRisk: A Lifecycle-Oriented Benchmark for Agent Harness Safety",
      "author_or_publisher": "HarnessRisk research team",
      "publication_or_revision_date": "2026-08-18",
      "url": "https://arxiv.org/abs/2608.17597",
      "evidence_class": "preprint",
      "scope_and_caveat": "Lifecycle safety evaluation, not certification.",
      "accessed": "2026-09-10"
    },
    {
      "id": "aces",
      "title": "Evaluating Skills, Not Just Agents: Agentic Continuous Evaluation of Skills",
      "author_or_publisher": "ACES research team",
      "publication_or_revision_date": "2026-08-20",
      "url": "https://arxiv.org/abs/2608.20614",
      "evidence_class": "preprint",
      "scope_and_caveat": "Paired live skill evaluation.",
      "accessed": "2026-09-10"
    },
    {
      "id": "constraints",
      "title": "Harness Engineering for Predictable Agentic Systems: An Empirical Study of Deterministic Execution Constraints",
      "author_or_publisher": "Saransh Dhage",
      "publication_or_revision_date": "2026-08-25",
      "url": "https://arxiv.org/html/2608.26197v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Mixed first-pass results; two synthetic linear tasks and two models limit generalization.",
      "accessed": "2026-09-10"
    },
    {
      "id": "frontier-launch",
      "title": "Introducing FrontierHarness Eval",
      "author_or_publisher": "Shilin Zhu and Shiqi Mei / Runta",
      "publication_or_revision_date": "2026-09-01",
      "url": "https://runta.com/blog/introducing-frontierharness-eval/",
      "evidence_class": "vendor benchmark report",
      "scope_and_caveat": "Reported setup, compatibility caveats and cost label; data audit finds label discrepancy and missing costs.",
      "accessed": "2026-09-10"
    },
    {
      "id": "harnessdev",
      "title": "HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?",
      "author_or_publisher": "HarnessDev research team / ByteDance Seed and collaborators",
      "publication_or_revision_date": "2026-09-01",
      "url": "https://arxiv.org/html/2609.01437v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Creation/evolution, held-out tasks, executor transfer, limited evolution trajectories.",
      "accessed": "2026-09-10"
    },
    {
      "id": "hoh",
      "title": "Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement",
      "author_or_publisher": "Haoyang Yan and colleagues",
      "publication_or_revision_date": "2026-09-01",
      "url": "https://arxiv.org/html/2609.01481v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "Outer workflow and continuation controls; not an independent reproduction.",
      "accessed": "2026-09-10"
    },
    {
      "id": "heart",
      "title": "Harness Engineering in LLM Tool Use via Agent-Native Reusable Tool Primitives",
      "author_or_publisher": "Haibo Jin, Suijin Wang, Xucheng Yu, Haojing Luo, Haohan Wang",
      "publication_or_revision_date": "2026-09-01",
      "url": "https://arxiv.org/html/2609.01736v1",
      "evidence_class": "preprint v1",
      "scope_and_caveat": "ToolFace and model-mediated tool interface; authorization implications are editorial analysis.",
      "accessed": "2026-09-10"
    },
    {
      "id": "frontier-data",
      "title": "FrontierHarness Eval public task-level results",
      "author_or_publisher": "FrontierHarness Eval contributors",
      "publication_or_revision_date": "2026-09-08",
      "url": "https://github.com/frontier-harness-eval/eval/blob/e837a70bd6beb4e72eeeda62dd06e3bd34f6cb63/results/eval-data.json",
      "evidence_class": "pinned dataset",
      "scope_and_caveat": "Original arithmetic independently recomputed; trials and invoices not independently verified.",
      "accessed": "2026-09-10"
    },
    {
      "id": "frontier-workflow",
      "title": "FrontierHarness Eval evaluation workflow",
      "author_or_publisher": "FrontierHarness Eval contributors",
      "publication_or_revision_date": null,
      "url": "https://github.com/frontier-harness-eval/eval",
      "evidence_class": "living repository",
      "scope_and_caveat": "New runs default non-comparable; baseline network allowlist not recorded.",
      "accessed": "2026-09-10"
    },
    {
      "id": "skills",
      "title": "SkillEvaluator",
      "author_or_publisher": "NVIDIA",
      "publication_or_revision_date": null,
      "url": "https://github.com/NVIDIA/SkillEvaluator",
      "evidence_class": "living repository",
      "scope_and_caveat": "Public multi-tier skill evaluation machinery; check version and support status.",
      "accessed": "2026-09-10"
    },
    {
      "id": "hoh-code",
      "title": "HarnessOfHarness",
      "author_or_publisher": "HoH authors",
      "publication_or_revision_date": null,
      "url": "https://github.com/Flesymeb/HarnessOfHarness",
      "evidence_class": "living project repository",
      "scope_and_caveat": "README explicitly describes HoH-lite as forthcoming at review date.",
      "accessed": "2026-09-10"
    },
    {
      "id": "self-developing",
      "title": "Self-Developing Agents",
      "author_or_publisher": "ByteDance Seed and TokenWave",
      "publication_or_revision_date": "2026-09-01",
      "url": "https://self-developing-agents.github.io/",
      "evidence_class": "research project page",
      "scope_and_caveat": "Research framing and paper links; not proof of a complete installable distribution.",
      "accessed": "2026-09-10"
    }
  ]
}