{
  "title": "The Handoff: AI Work Now Passes Between People, Programs and Other AIs. What Survives Each Pass Decides Whether It Holds Up",
  "slug": "the-handoff",
  "url": "https://emberandoakinstitute.com/blog/the-handoff.html",
  "date": "2026-11-16",
  "category": "AI news",
  "tags": [
    "artificial intelligence",
    "robotics",
    "AI agents",
    "physics",
    "AI safety",
    "AI news"
  ],
  "summary": "AI work is becoming a relay: physicist Nathaniel Craig credits GPT-6 Astra with leading the proof of a candidate lattice definition of the Standard Model, checked by three lattice experts and a Lean formalization; ACE Robotics' coding AI wrote frozen robot programs that succeeded 221/240 times (help on request 86.7% vs up front 53.3%); a CMU/USC study found a cheaper helper model let a planted fake record through on up to 75/180 questions versus about 1/180 when the strong model read itself.",
  "key_facts": [
    {
      "fact": "OpenAI published Nathaniel Craig's guest post on Nov. 15, 2026, the first guest post on its science blog; Craig leads UC Santa Barbara's particle theory group, is a permanent member of its KITP and a visiting scientist at OpenAI.",
      "source": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "fact": "Craig's Oct. 6, 2026 preprint proposes a candidate nonperturbative lattice definition of the Standard Model; 60+63 pages, 12 figures; mathematical companion and Lean formalization.",
      "source": "https://arxiv.org/abs/2610.09093"
    },
    {
      "fact": "Craig had the refinement strategy; GPT-6 Astra suggested how to carry fields to finer grids, showed convergence, and 'took the lead in developing the detailed proof'.",
      "source": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "fact": "Lattice field theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed Astra's proof and asked probing questions.",
      "source": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "fact": "Standard Model lacks a nonperturbative definition; question open for more than fifty years; phase prescription for full Standard Model open for more than 25 years after Lüscher's work on simpler theories.",
      "source": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "fact": "Craig: 'far more an answer in principle than an answer in practice'; 'Scientific dialogue among experts is required'; reflection positivity at finite lattice spacing still open.",
      "source": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "fact": "ACE-CaP-0 (arXiv Nov. 12, 2026): coding agent Claude Opus 5.5 writes/revises robot programs, can request kinesthetic demonstrations, run simulation, or test on robot; final program frozen and runs without a human or coding agent in the loop.",
      "source": "https://arxiv.org/abs/2611.16872"
    },
    {
      "fact": "Eight real-world tasks, 30 trials each: overall 92.1% (221/240); Charger Insertion 30/30; Bread Packing 25/30; Chicken Plating 29/30; mean execution 0.97–2.27 minutes.",
      "source": "https://arxiv.org/abs/2611.16872"
    },
    {
      "fact": "Charger insertion guidance: robot demonstrations 86.7% (26/30), robot-egocentric human video 73.3% (22/30), textual guidance 60%; on-request 86.7% vs upfront 53.3% (16/30); development time 3.7 vs 7.2 hours.",
      "source": "https://arxiv.org/abs/2611.16872"
    },
    {
      "fact": "Wang & Mannby (CMU/USC/Prosaic Intelligence), arXiv Nov. 12, 2026, accepted to NeurIPS 2026 Meta-Agents workshop: one forged file among 12 genuine; 180 fictional facts; strong orchestrators reading directly made 2 harmful revisions in 1,080 trials.",
      "source": "https://arxiv.org/abs/2611.17520"
    },
    {
      "fact": "Delegating to weaker sub-agent: Claude Code (Opus 5→Haiku 4.5) 75/180 stacked, 20/180 gloss; Codex (GPT-5.6 Sol→Terra) 44/180 gloss; same-tier sub-agent 0/180, 0/180, 1/180.",
      "source": "https://arxiv.org/abs/2611.17520"
    },
    {
      "fact": "Orchestrator opening forged file named missing provenance 180/180; weaker sub-agent summary carried it 96/180; grounded verifier with source record: aggregators 0/780 wrong.",
      "source": "https://arxiv.org/abs/2611.17520"
    },
    {
      "fact": "Authors: 'An agentic system is only as safe as its weakest link.'",
      "source": "https://arxiv.org/abs/2611.17520"
    }
  ],
  "sources": [
    {
      "name": "Nathaniel Craig, “A step toward defining the Standard Model on a lattice,” OpenAI science blog guest post (Nov. 15, 2026)",
      "url": "https://openai.com/index/step-toward-defining-the-standard-model/"
    },
    {
      "name": "Nathaniel Craig, “A gauge-invariant measure for lattice chiral fermions,” arXiv:2610.09093 (Oct. 6, 2026)",
      "url": "https://arxiv.org/abs/2610.09093"
    },
    {
      "name": "ACE Robotics CaP Team, “ACE-CaP-0: Real-World Code-as-Policies in the Era of Experience,” arXiv:2611.16872 (Nov. 12, 2026)",
      "url": "https://arxiv.org/abs/2611.16872"
    },
    {
      "name": "ACE Robotics, ACE-CaP-0 project page (Nov. 2026)",
      "url": "https://acerobotics2025.github.io/ACE-CaP/"
    },
    {
      "name": "Keegan Wang & Anantika Mannby, “Only as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-Agent,” arXiv:2611.17520 (Nov. 12, 2026; NeurIPS 2026 Meta-Agents workshop)",
      "url": "https://arxiv.org/abs/2611.17520"
    }
  ]
}