{
  "title": "Harness engineering: September 15 research extension",
  "review_date": "2026-09-15",
  "audience": "Practicing AI engineers and technical founders",
  "scope": "Selective primary-source synthesis covering execution, evaluation, optimization, context delegation, safety, and managed runtimes; extends the September 10 series.",
  "method": "Primary pages, selected paper full texts, official documentation, and repository surfaces reviewed. Historical and recent sources distinguished. No new model trials, hosted-service integration tests, or independent reproduction of reported benchmark results.",
  "limitations": [
    "Coverage is selective, not a systematic exhaustive literature census.",
    "Repository visibility is not installability or independently verified performance.",
    "Proposed experiments and example packets are not ContextOS normative runtime schemas.",
    "Source dates follow the primary records; the 2609.00006 identifier and stated July submission date are explicitly retained."
  ],
  "articles": [
    "/blog/harness-engineering-september-15-field-guide",
    "/blog/subagent-context-inheritance-independent-review",
    "/blog/managed-agent-harness-control-boundaries",
    "/blog/harness-engineering-research-publishing-roadmap"
  ],
  "sources": [
    {
      "id": "swe-agent",
      "title": "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering",
      "author_or_publisher": "John Yang and colleagues",
      "publication_or_revision_date": "2024-11-11 (v3; first submitted 2024-05-06)",
      "url": "https://arxiv.org/abs/2405.15793",
      "evidence_class": "research paper",
      "scope_and_caveat": "Historical interface-design evidence; no current leaderboard comparison inferred.",
      "accessed": "2026-09-15"
    },
    {
      "id": "camel",
      "title": "Defeating Prompt Injections by Design",
      "author_or_publisher": "Edoardo Debenedetti and colleagues",
      "publication_or_revision_date": "2025-06-24 (v2)",
      "url": "https://arxiv.org/abs/2503.18813",
      "evidence_class": "research paper",
      "scope_and_caveat": "Control and data-flow separation; guarantees are conditional on the specified architecture and threat model.",
      "accessed": "2026-09-15"
    },
    {
      "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": "Durable progress and incremental sessions; workload-specific experience.",
      "accessed": "2026-09-15"
    },
    {
      "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, and evaluation in application development; not a universal topology.",
      "accessed": "2026-09-15"
    },
    {
      "id": "trace-guided",
      "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": "Trace-guided fixed-model intervention; results not independently rerun.",
      "accessed": "2026-09-15"
    },
    {
      "id": "infrastructure",
      "title": "Quantifying infrastructure noise in agentic coding evals",
      "author_or_publisher": "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 configuration changes evaluation conditions.",
      "accessed": "2026-09-15"
    },
    {
      "id": "anatomy",
      "title": "Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents — A Source-Code Study of Eleven Systems",
      "author_or_publisher": "Paul Barbaste, Tristan Darrigol, Germain Vu, Tom Wiltberger",
      "publication_or_revision_date": "2026-07-15 as stated by arXiv submission record; v1 identifier 2609.00006",
      "url": "https://arxiv.org/html/2609.00006v1",
      "evidence_class": "preprint / observational source-code study",
      "scope_and_caveat": "Dates preserved from primary record despite identifier mismatch. Corpus observations do not establish causal superiority or universal absences.",
      "accessed": "2026-09-15"
    },
    {
      "id": "risk",
      "title": "HarnessRisk: A Lifecycle-Oriented Benchmark for Agent Harness Safety",
      "author_or_publisher": "Yajing Bai and colleagues",
      "publication_or_revision_date": "2026-08-18 (v1)",
      "url": "https://arxiv.org/abs/2608.17597",
      "evidence_class": "preprint benchmark",
      "scope_and_caveat": "Lifecycle safety and difference between recognition and prevention; sandbox results are not production prevalence.",
      "accessed": "2026-09-15"
    },
    {
      "id": "harnessdev",
      "title": "HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?",
      "author_or_publisher": "Yuhao Wu and colleagues",
      "publication_or_revision_date": "2026-09-01 (v1)",
      "url": "https://arxiv.org/abs/2609.01437",
      "evidence_class": "preprint benchmark",
      "scope_and_caveat": "Creation and evolution are distinct; transfer is conditional on domain and executor.",
      "accessed": "2026-09-15"
    },
    {
      "id": "evosafe",
      "title": "EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents",
      "author_or_publisher": "Nanxi Li and colleagues",
      "publication_or_revision_date": "2026-09-05 (v1)",
      "url": "https://arxiv.org/html/2609.05903v1",
      "evidence_class": "preprint experiment",
      "scope_and_caveat": "Reported safety/utility outcomes; residual failures remain. Fresh-context review motivates a separate proposed study.",
      "accessed": "2026-09-15"
    },
    {
      "id": "same-model",
      "title": "Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite",
      "author_or_publisher": "Mohsen Arjmandi",
      "publication_or_revision_date": "2026-09-08 revision as stated in v1",
      "url": "https://arxiv.org/html/2609.11987v1",
      "evidence_class": "preprint experiment with corrected telemetry",
      "scope_and_caveat": "Unresolved mean advantage, private tasks, post-hoc workload split, and incomplete billed-cost coverage. No independent reproduction claimed.",
      "accessed": "2026-09-15"
    },
    {
      "id": "context-modes",
      "title": "Organizing Context in a Multi-Agent Harness",
      "author_or_publisher": "Thushanth Bengre and Chester Curme / LangChain",
      "publication_or_revision_date": "2026-09-08",
      "url": "https://www.langchain.com/blog/organizing-context-in-a-multi-agent-harness",
      "evidence_class": "first-party implementation report",
      "scope_and_caveat": "Forked and isolated contexts; performance depends on workload and caching.",
      "accessed": "2026-09-15"
    },
    {
      "id": "connections",
      "title": "Connections: Managed credentials and per-caller identity for Managed Deep Agents",
      "author_or_publisher": "Victor Moreira / LangChain",
      "publication_or_revision_date": "2026-09-09",
      "url": "https://www.langchain.com/blog/connections-managed-credentials-and-per-caller-identity-for-managed-deep-agents",
      "evidence_class": "first-party prerelease feature report",
      "scope_and_caveat": "v0.7.0+; ownership and credential type are separate. Identity does not imply task-specific authorization.",
      "accessed": "2026-09-15"
    },
    {
      "id": "agents-api",
      "title": "Agents API overview",
      "author_or_publisher": "OpenAI",
      "publication_or_revision_date": "living documentation reviewed 2026-09-15",
      "url": "https://developers.openai.com/api/docs/guides/agents-api/overview",
      "evidence_class": "official product documentation",
      "scope_and_caveat": "Managed sessions and execution facilities; account access and operational performance not tested.",
      "accessed": "2026-09-15"
    },
    {
      "id": "agents-api-launch",
      "title": "Introducing the Agents API",
      "author_or_publisher": "OpenAI",
      "publication_or_revision_date": "2026-09-10",
      "url": "https://openai.com/index/introducing-the-agents-api/",
      "evidence_class": "first-party launch announcement",
      "scope_and_caveat": "Source of launch date; customer testimonials not used as comparative evidence.",
      "accessed": "2026-09-15"
    },
    {
      "id": "managed-architecture",
      "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": "Session, harness, sandbox separation and model-dependent workaround removal.",
      "accessed": "2026-09-15"
    },
    {
      "id": "deepagents-code",
      "title": "Deep Agents",
      "author_or_publisher": "LangChain",
      "publication_or_revision_date": "living repository reviewed 2026-09-15",
      "url": "https://github.com/langchain-ai/deepagents",
      "evidence_class": "public source repository",
      "scope_and_caveat": "Availability and documentation inspected; no pinned checkout or local execution claimed.",
      "accessed": "2026-09-15"
    },
    {
      "id": "subagent-docs",
      "title": "Subagents",
      "author_or_publisher": "LangChain",
      "publication_or_revision_date": "living documentation reviewed 2026-09-15",
      "url": "https://docs.langchain.com/oss/python/deepagents/subagents",
      "evidence_class": "official implementation documentation",
      "scope_and_caveat": "Check language-specific versions and defaults before adoption.",
      "accessed": "2026-09-15"
    },
    {
      "id": "evosafe-code",
      "title": "EvoSafeHarness official implementation",
      "author_or_publisher": "SaFo-Lab",
      "publication_or_revision_date": "living repository reviewed 2026-09-15",
      "url": "https://github.com/SaFo-Lab/EvoSafeHarness",
      "evidence_class": "public research repository",
      "scope_and_caveat": "Public implementation surface inspected; dependency installation and results not reproduced.",
      "accessed": "2026-09-15"
    },
    {
      "id": "hoh-code",
      "title": "HarnessOfHarness",
      "author_or_publisher": "Flesymeb / HoH authors",
      "publication_or_revision_date": "living repository reviewed 2026-09-15",
      "url": "https://github.com/Flesymeb/HarnessOfHarness",
      "evidence_class": "public project repository",
      "scope_and_caveat": "README still announces HoH-lite as coming soon. Project artifacts do not establish a complete lightweight runtime release.",
      "accessed": "2026-09-15"
    }
  ]
}