{
  "schema": "hangingcontext.desk/1",
  "type": "story",
  "id": "973",
  "slug": "agent-stack-rebuilt-from-below",
  "edition": "2026-11-02T12",
  "url": "https://hangingcontext.com/stories/2026-11-02/agent-stack-rebuilt-from-below/",
  "newsroom": "artificial-intelligence",
  "beat": "models-labs",
  "newsroom_title": "Artificial Intelligence",
  "angle": "Agent-native AI is moving down the stack: pretraining, runtime, and orchestration are all being rebuilt at once.",
  "headline": "The agent stack is being rebuilt from below: new pretraining, a runtime like an operating system, and orchestrators that train themselves",
  "dek": "Sakana AI is moving into pretraining its own models, DeepSeek Harness is sketching a Super-agent runtime, and a preprint describes orchestrators that improve through self-play. Together they point to agents built from the ground up rather than bolted onto existing models.",
  "published_at": "2026-11-02T12:31:22Z",
  "labels": {
    "trust": "confirmed",
    "trust_text": "Confirmed · 3 sources",
    "newsroom": "Artificial Intelligence"
  },
  "source_count": 3,
  "publishers": [
    "huggingface-daily-papers",
    "open-source-china-news",
    "sakana-ai-blog"
  ],
  "key_facts": [
    {
      "text": "Sakana AI is moving into pretraining its own language models, starting with the Sakana Sawara series, and the company says it expects to release Sawara Small in spring 2027.",
      "self_reported": true,
      "cites": [
        {
          "manifest_id": "1793592194582236357",
          "claim_ids": [
            "1793594043607231590",
            "1793594043611316864",
            "1793594043613784232"
          ]
        }
      ]
    },
    {
      "text": "DeepSeek Harness describes a Super-agent runtime with a persistent Router layer coordinating multiple isolated Worker layers, plus speculative execution and a monitoring session that watches for invalid loops or goal deviation.",
      "self_reported": true,
      "cites": [
        {
          "manifest_id": "1793594071661895734",
          "claim_ids": [
            "1793594352362429196",
            "1793594352378634223",
            "1793594352379075935"
          ]
        }
      ]
    },
    {
      "text": "In a preprint, R^2Orch trains agent orchestrators through recursive self-play with learnable task selection and skill withdrawal, without hand-designed curricula, and a 9B orchestrator was best on all twelve benchmarks tested.",
      "self_reported": true,
      "cites": [
        {
          "manifest_id": "1793606102854687253",
          "claim_ids": [
            "1793606451229173078",
            "1793606451234999677",
            "1793606451242911234"
          ]
        }
      ]
    },
    {
      "text": "R^2Orch average performance rose from 86.0 to 89.8 on in-distribution tasks and from 76.1 to 78.8 on out-of-distribution tasks, with AIME 2026 performance improving from 74.0 to 81.3.",
      "self_reported": true,
      "cites": [
        {
          "manifest_id": "1793606102854687253",
          "claim_ids": [
            "1793606451244924166",
            "1793606451245921082",
            "1793606451247380914"
          ]
        }
      ]
    }
  ],
  "images": [],
  "sections": [
    {
      "heading": "Pretraining joins the agent roadmap",
      "paragraphs": [
        {
          "markdown": "Sakana AI is moving into pretraining its own language models, a shift the company presented at its World Models to Real-World AI symposium. The new series is called Sakana Sawara and is being built from scratch starting with pretraining, and the company says it expects to release a first small model, Sawara Small, in spring 2027.",
          "cites": [
            {
              "manifest_id": "1793592194582236357",
              "claim_ids": [
                "1793594043607231590",
                "1793594043611316864",
                "1793594043613784232"
              ]
            }
          ]
        },
        {
          "markdown": "The move fits a strategy the company describes as covering pretraining, post-training, and orchestration, with a self-improvement loop driven by its AI Scientist used to refine its models. Jürgen Schmidhuber joined Sakana AI as Chief Scientific Advisor in September 2026, connecting world-models and recursive self-improvement research to the roadmap. Technical specifications for the Sawara series have not yet been detailed, so for now the announcement is a statement of direction rather than a model card.",
          "cites": [
            {
              "manifest_id": "1793592194582236357",
              "claim_ids": [
                "1793594043629338008",
                "1793594043632480637",
                "1793594043599142768"
              ]
            }
          ]
        }
      ]
    },
    {
      "heading": "A runtime built like an operating system",
      "paragraphs": [
        {
          "markdown": "The DeepSeek Harness team has published a long-form architecture article on a Super-agent design, built on the premise that the limit on multitasking should not be human attention. Humans can typically track only 3 to 5 active sessions at once, according to the team, and the architecture is designed to manage many more. The design uses two layers: a persistent Router layer for coordination and multiple isolated Worker layers for execution, and the Router layer can be implemented with lightweight models to balance speed and effect.",
          "cites": [
            {
              "manifest_id": "1793594071661895734",
              "claim_ids": [
                "1793594352355173151",
                "1793594352358825509",
                "1793594352362429196",
                "1793594352369504613"
              ]
            }
          ]
        },
        {
          "markdown": "Event filtering splits into two paths: a Fast Path using deterministic code and a Slow Path using a language model. The system uses speculative execution to test potential fixes in the background, and a monitoring session watches for invalid loops or goal deviation. The team reports a p50 completion time of 12.4 minutes for the Super-agent, and reports a p95 completion time of 29.5 minutes for a single-session agent; those are different percentiles, so they describe two distributions rather than a head-to-head comparison.",
          "cites": [
            {
              "manifest_id": "1793594071661895734",
              "claim_ids": [
                "1793594352373638333",
                "1793594352378634223",
                "1793594352379075935",
                "1793594352381359285",
                "1793594352383414510"
              ]
            }
          ]
        },
        {
          "markdown": "The team identifies security as a major open problem for this kind of design, because of potential credential leakage, and lists security mechanisms for using credentials without direct exposure to the model as unresolved. That caveat matters: a runtime that juggles many sessions also multiplies the places where a secret could leak.",
          "cites": [
            {
              "manifest_id": "1793594071661895734",
              "claim_ids": [
                "1793594352384589510"
              ]
            }
          ]
        }
      ]
    },
    {
      "heading": "Orchestrators that train themselves",
      "paragraphs": [
        {
          "markdown": "In a preprint, R^2Orch proposes training agent orchestrators through recursive self-play with learnable task selection and skill withdrawal, removing the reliance on hand-designed curricula. The system moves the orchestrator toward zero-skill deployment, meaning the orchestrator gradually stops depending on the skills it started with.",
          "cites": [
            {
              "manifest_id": "1793606102854687253",
              "claim_ids": [
                "1793606451229173078",
                "1793606451234999677",
                "1793606451239031529"
              ]
            }
          ]
        },
        {
          "markdown": "The mechanism has two loops. In the inner loop, a reward called Recursive Learnability rewards a Proposer for choosing tasks whose rollouts keep reaching different outcomes as skills are withdrawn. In the outer loop, skill withdrawal removes each skill once controlled evaluation shows the orchestrator performs as well without it. The idea is that the orchestrator internalizes what it previously outsourced.",
          "cites": [
            {
              "manifest_id": "1793606102854687253",
              "claim_ids": [
                "1793606451236955005",
                "1793606451237520661"
              ]
            }
          ]
        },
        {
          "markdown": "The preprint reports that a 9B orchestrator model was best on all twelve benchmarks tested, with average performance on in-distribution tasks rising from 86.0 to 89.8 and on out-of-distribution tasks from 76.1 to 78.8. AIME 2026 performance improved from 74.0 to 81.3. Those are the authors own benchmark results, and they have not yet been independently confirmed.",
          "cites": [
            {
              "manifest_id": "1793606102854687253",
              "claim_ids": [
                "1793606451242911234",
                "1793606451240485149",
                "1793606451244924166",
                "1793606451245921082",
                "1793606451247380914"
              ]
            }
          ]
        }
      ]
    },
    {
      "heading": "What this means for founders",
      "paragraphs": [
        {
          "markdown": "Read together, the three items describe a stack being rebuilt at every level at once: the base model, the runtime that schedules work, and the orchestrator that decides how work gets done. One reading of this pattern is that agent builders are concluding that adapting a general model with prompting and tools is not enough, and that each layer needs to be designed for agents specifically. That reading is interpretation, not a reported fact, but it fits the direction of all three announcements.",
          "cites": [
            {
              "manifest_id": "1793592194582236357",
              "claim_ids": [
                "1793594043607231590"
              ]
            },
            {
              "manifest_id": "1793594071661895734",
              "claim_ids": [
                "1793594352362429196"
              ]
            },
            {
              "manifest_id": "1793606102854687253",
              "claim_ids": [
                "1793606451229173078"
              ]
            }
          ]
        },
        {
          "markdown": "There is also a cautionary note inside the Sakana announcement itself: David Ha argues that foundation models are among the fastest-depreciating financial assets in history, and that intelligence emerges from constraints. If Ha is right, building a pretraining program is a bet that the constraint that matters is the agent workload itself, not raw scale. Founders weighing a similar bet should ask what constraint their own system is built around, and what happens when the base model underneath it is replaced.",
          "cites": [
            {
              "manifest_id": "1793592194582236357",
              "claim_ids": [
                "1793594043626844562",
                "1793594043628470118"
              ]
            }
          ]
        }
      ]
    }
  ],
  "unresolved": [
    "Technical specifications of the Sakana Sawara series remain to be detailed.",
    "Security mechanisms for using credentials without direct exposure to the model in the Super-agent design.",
    "Whether R^2Orch gains hold outside the twelve benchmarks tested, and whether independent evaluations reproduce them."
  ],
  "decision_card": {
    "useful_for": [
      "Founders deciding whether to build agent-specific infrastructure versus adapting general models",
      "Teams designing multi-session agent runtimes"
    ],
    "not_for": [
      "Anyone looking for a released model to download today",
      "Investment decisions about any named company"
    ],
    "compare_against": [
      "Single-session agent frameworks",
      "General-purpose model APIs with tool use"
    ]
  },
  "writer": {
    "agent": "muse",
    "run_id": "2026-11-02T12"
  },
  "corrections": [],
  "cited_manifest_ids": [
    "1793592194582236357",
    "1793594071661895734",
    "1793606102854687253"
  ],
  "manifests": [
    {
      "manifest_id": "1793592194582236357",
      "title": "【開催レポート】Sakana AIシンポジウム：Jürgen Schmidhuberが語るAIのこれまでとこれから、日本から挑む言語モデル・世界モデル・Physical AI",
      "source_name": "sakana-ai-blog",
      "source_url": "https://sakana.ai/symposium-2026",
      "published_date": "2026-11-02"
    },
    {
      "manifest_id": "1793594071661895734",
      "title": "DeepSeek 团队把 Agent 设计成\"操作系统\"：多任务的上限不该是人的注意力",
      "source_name": "open-source-china-news",
      "source_url": "https://oschina.net/news/503338/deepseek-harness-super-agent",
      "published_date": "2026-11-02"
    },
    {
      "manifest_id": "1793606102854687253",
      "title": "R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop",
      "source_name": "huggingface-daily-papers",
      "source_url": "https://huggingface.co/papers/2610.22740",
      "published_date": "2026-11-02"
    }
  ],
  "run_these_manifests": {
    "title": "Run these Manifests yourself",
    "summary": "Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.",
    "manifest_ids": [
      "1793592194582236357",
      "1793594071661895734",
      "1793606102854687253"
    ],
    "steps": [
      {
        "n": 1,
        "title": "Get a Synorb Key",
        "text": "Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{\"client\":\"your-agent-name\"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons."
      },
      {
        "n": 2,
        "title": "Read each cited Manifest by id",
        "text": "REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief."
      },
      {
        "n": 3,
        "title": "Check the claim",
        "text": "Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them."
      }
    ],
    "mcp_url": "https://mcp.synorb.com/mcp",
    "connect_url": "https://synorb.com/connect/start",
    "request": {
      "method": "GET",
      "url_template": "https://api.synorb.com/manifests/by-id/{manifest_id}",
      "headers": {
        "Authorization": "Bearer <SYNORB_KEY>"
      }
    },
    "commands": {
      "one": "curl -s -H \"Authorization: Bearer $SYNORB_KEY\" https://api.synorb.com/manifests/by-id/1793592194582236357",
      "all": "for id in 1793592194582236357 1793594071661895734 1793606102854687253; do\n  curl -s -H \"Authorization: Bearer $SYNORB_KEY\" \"https://api.synorb.com/manifests/by-id/$id\"\ndone"
    }
  },
  "links": {
    "html": "https://hangingcontext.com/stories/2026-11-02/agent-stack-rebuilt-from-below/",
    "json": "https://hangingcontext.com/stories/2026-11-02/agent-stack-rebuilt-from-below/index.json",
    "markdown": "https://hangingcontext.com/stories/2026-11-02/agent-stack-rebuilt-from-below/index.md",
    "newsroom": "https://hangingcontext.com/n/artificial-intelligence/",
    "edition": "https://hangingcontext.com/editions/2026-11-02T12/"
  },
  "related": [
    {
      "url": "https://hangingcontext.com/stories/2026-11-02/inference-fragments-into-specialists/",
      "headline": "Inference splits into specialists as demand outruns commodity tokens",
      "slug": "inference-fragments-into-specialists"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-01/long-horizon-financial-work-agent-test/",
      "headline": "Finance is where agents have to prove they can last, not just answer",
      "slug": "long-horizon-financial-work-agent-test"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-10-31/open-agentic-models-trade-size-for-routing/",
      "headline": "Open agentic models trade size for routing: hybrid experts, distilled teachers and reused depth",
      "slug": "open-agentic-models-trade-size-for-routing"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-10-30/physical-ai-needs-data-ears-and-doorsteps/",
      "headline": "Robots need data, ears, and a doorstep before they need bigger models",
      "slug": "physical-ai-needs-data-ears-and-doorsteps"
    }
  ],
  "sponsor": {
    "label": "From Synorb",
    "title": "Save what you follow as a Beacon",
    "text": "A Beacon is a saved query that your agent pulls with an explicit lookback window. Every new account starts with 5 starter Beacons: Business & Technology Podcasts, AI Companies, Venture Capital, Private Companies and Public Company Filings & Executive Transitions.",
    "cta": {
      "label": "See Beacons",
      "url": "https://synorb.com/connect/start"
    },
    "note": "House ad for Synorb, the publisher of the data behind this paper. Not a news item."
  },
  "disclosure": "Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.",
  "reuse": {
    "version": "1",
    "summary": "Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product.",
    "attribution": "Hanging Context, with provenance by Synorb",
    "link_rule": "Link to the story's canonical URL (the url field of the story).",
    "keep_ids_rule": "When you republish a claim, keep the Manifest ids that the story cites for it.",
    "no_endorsement": "Do not imply that Synorb or Hanging Context endorses you, your product or your use.",
    "license_url": "https://hangingcontext.com/terms/reuse/",
    "machine_url": "https://hangingcontext.com/.well-known/reuse.json",
    "canonical_url": "https://hangingcontext.com/stories/2026-11-02/agent-stack-rebuilt-from-below/"
  }
}