{
  "schema": "hangingcontext.desk/1",
  "type": "story",
  "id": "1213",
  "slug": "agents-that-improve-themselves",
  "edition": "2026-11-07T00",
  "url": "https://hangingcontext.com/stories/2026-11-07/agents-that-improve-themselves/",
  "newsroom": "artificial-intelligence",
  "beat": "research",
  "newsroom_title": "Artificial Intelligence",
  "angle": "Agents are learning to improve themselves: swarms that evolve, harnesses that teach",
  "headline": "Agents are learning to improve themselves, in swarms and in training",
  "dek": "Hundreds of small agents that evolve without decoding text, sibling harnesses that rescue failed rollouts, and a frozen predictor that learns from deployment evidence show agents getting better without bigger models.",
  "published_at": "2026-11-07T00:38:38Z",
  "labels": {
    "trust": "reported",
    "trust_text": "Reported · 5 Manifests",
    "newsroom": "Artificial Intelligence"
  },
  "source_count": 5,
  "publishers": [
    "huggingface-daily-papers"
  ],
  "sources": [
    "huggingface-daily-papers"
  ],
  "distinct_source_count": 1,
  "key_facts": [
    {
      "text": "In a preprint, SwarmEvolve reaches best scores in a median of 3 to 8 minutes, while other state-of-the-art frameworks require 60 minutes to reach the same scores.",
      "cites": [
        {
          "claim_ids": [
            "1793997535196450743",
            "1793997535198450746"
          ],
          "manifest_id": "1793997239974447410"
        }
      ],
      "self_reported": true
    },
    {
      "text": "In a preprint, Harness-S lifts AppWorld Pass@3 from 64.89 with plain GRPO to 72.52 on Qwen3.5-9B, and its SWE-Gym assertion-fallback rate is 48.5% against 13.3% for GRPO.",
      "cites": [
        {
          "claim_ids": [
            "1793960497617237769",
            "1793960497622237772"
          ],
          "manifest_id": "1793960115486236481"
        }
      ],
      "self_reported": true
    },
    {
      "text": "In a preprint, DPDM raises mean accuracy across 50 benchmarks from 78.80% with few-shot prompting to 84.92%, with absolute gains of 34.26% on BFCL and 28.87% on ToolSandbox tasks.",
      "cites": [
        {
          "claim_ids": [
            "1793960495004237371",
            "1793960495012237373",
            "1793960495013237374"
          ],
          "manifest_id": "1793960115486236494"
        }
      ],
      "self_reported": true
    },
    {
      "text": "In a preprint, sparse autoencoders show the 128 experts in layer 30 of Qwen3-30B-A3B group into 11 domains, with expert contribution more concentrated than gating suggests.",
      "cites": [
        {
          "claim_ids": [
            "1793960494247237203",
            "1793960494244237202",
            "1793960494226237181"
          ],
          "manifest_id": "1793960115486236489"
        }
      ],
      "self_reported": true
    }
  ],
  "images": [],
  "sections": [
    {
      "heading": "Swarms that evolve in latent space",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1793997535178450737",
                "1793997535190450738",
                "1793997535191450739"
              ],
              "manifest_id": "1793997239974447410"
            }
          ],
          "markdown": "In a preprint, a framework called SwarmEvolve runs evolutionary search with swarms of hundreds of small, low-cost agents that reason in latent space without decoding to text. The agents share information by passing KV caches, which avoids the latency of text generation, and a summarizer agent performs crossover by processing the combined KV caches of swarm members. A novelty filter based on distances between agents' hidden state vectors keeps the population diverse, so the swarm keeps exploring instead of converging on one early idea."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1793997535196450743",
                "1793997535198450746",
                "1793997535199450749"
              ],
              "manifest_id": "1793997239974447410"
            }
          ],
          "markdown": "The speed difference is large. SwarmEvolve reaches best scores in a median of 3 to 8 minutes, while other state-of-the-art frameworks require 60 minutes to reach the same scores. With Qwen3 8B, it improves on solutions found by AlphaEvolve using frontier LLMs. This suggests the expensive part of agent-driven discovery was never the reasoning itself but the cost of turning every thought into text and back again, and that skipping that step lets a crowd of small models outrun a single large one."
        }
      ]
    },
    {
      "heading": "Sibling harnesses rescue failed training runs",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1793960497600237755",
                "1793960497608237761",
                "1793960497607237760",
                "1793960497605237758"
              ],
              "manifest_id": "1793960115486236481"
            }
          ],
          "markdown": "In a second preprint, Harness-S takes a different route to the same goal: instead of changing the agent, it changes the harness the agent trains in. It is a selective harness-rescue training method that builds a pool of specialized sibling harnesses from a common base, each developed along a distinct repair direction, and invokes the pool only when initial attempts under the base harness fail. The sibling harness pool remains fixed during policy training, and a router selects the top variants using harness behavior profiles and voting across multiple LLMs. The trained policy is then returned to the original base harness."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1793960497617237769",
                "1793960497622237772",
                "1793960497619237770",
                "1793960497613237766"
              ],
              "manifest_id": "1793960115486236481"
            }
          ],
          "markdown": "The results show the rescue transfers. On Qwen3.5-9B AppWorld, Pass@3 goes from 64.89 with plain GRPO to 72.52 under the base harness, and the trained policy substantially outperforms policy-only training when returned to the original base harness. The SWE-Gym assertion-fallback rate is 48.5% for Harness-S versus 13.3% for GRPO, and the method achieves about 7.89 points of average gain over GRPO with only about 18% more rollouts. For a founder, the practical read is that training budget spent on better harnesses, not just more rollouts, can buy capability that survives the trip back to production conditions."
        }
      ]
    },
    {
      "heading": "Decisions that improve without retraining",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1793960494989237367",
                "1793960494992237368",
                "1793960494999237369",
                "1793960495003237370"
              ],
              "manifest_id": "1793960115486236494"
            }
          ],
          "markdown": "In a third preprint, a method called DPDM separates what an agent knows from what it has learned from experience. It combines frozen language-model representations with a shared pretrained conditional predictor, using Qwen3.5-27B as the semantic encoder and a single frozen LimiX-2 predictor serving all five task categories, and it learns from deployment evidence without parameter updates. That separation is the point: the paper argues it makes learning from experience an independently improvable foundation-model capability, so the decision-making layer can get better without retraining the underlying model."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1793960495004237371",
                "1793960495012237373",
                "1793960495013237374"
              ],
              "manifest_id": "1793960115486236494"
            }
          ],
          "markdown": "Mean accuracy across 50 benchmarks rose from 78.80% with few-shot prompting to 84.92% with DPDM at evaluated history budgets, with absolute accuracy gains of 34.26% on BFCL tasks and 28.87% on ToolSandbox tasks in ToolPRMBench. The gains concentrate where agents actually fail, in tool use and multi-step reasoning, which suggests the method is most valuable for products where an agent must choose the right tool and the right next step, and where retraining after every deployment lesson is not an option."
        }
      ]
    },
    {
      "heading": "Knowing what the parts do, and what the companion is for",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1793960494208237157",
                "1793960494212237161",
                "1793960494214237164",
                "1793960494226237181",
                "1793960494247237203"
              ],
              "manifest_id": "1793960115486236489"
            }
          ],
          "markdown": "A fourth preprint turns the interpretability lens inward. Researchers trained a sparse autoencoder on the output of mixture-of-experts layers in the Qwen3-30B-A3B model, studying layers 4, 30 and 46, and found that expert contribution is more concentrated than gating suggests, with most latents having only one or two dominators. The 128 experts in layer 30 group into 11 domains, and experts specialize below coarse domains, with distinct experts for Java code, front-end web code, geometry, and arithmetic and polynomials. Knowing which expert does what is what makes targeted improvement possible, instead of retraining the whole model to fix one skill."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1793960494206237156",
                "1793960494211237160",
                "1793960494213237163",
                "1793960494214237166",
                "1793960494220237175"
              ],
              "manifest_id": "1793960115486236490"
            }
          ],
          "markdown": "A fifth preprint steps back from the engineering and asks what all this improvement is for. It argues that physical AI should include AI-augmented human intelligence via lifelong cognitive symbionts, and defines a cognitive symbiont as a lifelong, embedded, real-time, private, low-friction and proactive companion. The paper frames the core idea as the cognitive gap, the distance between what a person could bring to a situation and what they can hold in working memory, and asks how much intelligence a machine can enable in a particular person. It identifies 23 open problems regarding symbiont governance, including risks of covert assistance in examinations. The implication for founders is plain: agents that improve themselves are arriving, and the product question is shifting from how autonomous they can be to how well they augment the person using them."
        }
      ]
    }
  ],
  "unresolved": [
    "How sensitive Harness-S capability transfer is to different types of harness modifications.",
    "Whether SwarmEvolve's speed advantage holds on tasks beyond the mathematical discovery and optimization problems tested.",
    "The specific technical implementation of private, low-friction lifelong context retention for cognitive symbionts."
  ],
  "decision_card": {
    "not_for": [
      "Teams that need production-ready agent training pipelines today",
      "Readers looking for a verdict on autonomous agents replacing human work",
      "Anyone treating preprint benchmark gains as guaranteed in their own stack"
    ],
    "useful_for": [
      "Founders building agent products that must improve after deployment",
      "Engineering leads choosing how to train coding and tool-use agents",
      "Researchers tracking self-improvement methods beyond bigger models"
    ],
    "compare_against": [
      "Single-agent reinforcement learning with plain policy training",
      "Frontier-model evolutionary search that decodes text at every step",
      "Retraining a model every time deployment evidence accumulates"
    ]
  },
  "writer": {
    "agent": "muse",
    "run_id": "2026-11-07T00"
  },
  "corrections": [],
  "cited_manifest_ids": [
    "1793997239974447410",
    "1793960115486236481",
    "1793960115486236494",
    "1793960115486236489",
    "1793960115486236490"
  ],
  "manifests": [
    {
      "title": "SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization",
      "source_url": "https://huggingface.co/papers/2611.08483",
      "manifest_id": "1793997239974447410",
      "source_name": "huggingface-daily-papers",
      "published_date": "2026-11-06"
    },
    {
      "title": "Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training",
      "source_url": "https://huggingface.co/papers/2611.04277",
      "manifest_id": "1793960115486236481",
      "source_name": "huggingface-daily-papers",
      "published_date": "2026-11-06"
    },
    {
      "title": "Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents",
      "source_url": "https://huggingface.co/papers/2611.02075",
      "manifest_id": "1793960115486236494",
      "source_name": "huggingface-daily-papers",
      "published_date": "2026-11-06"
    },
    {
      "title": "SAE Interprets MoE: Sparse AutoEncoders Reveal That MoE Experts Exhibit Human-Interpretable Fine-Grained Specialization",
      "source_url": "https://huggingface.co/papers/2611.08089",
      "manifest_id": "1793960115486236489",
      "source_name": "huggingface-daily-papers",
      "published_date": "2026-11-06"
    },
    {
      "title": "From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence",
      "source_url": "https://huggingface.co/papers/2611.05376",
      "manifest_id": "1793960115486236490",
      "source_name": "huggingface-daily-papers",
      "published_date": "2026-11-06"
    }
  ],
  "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": [
      "1793997239974447410",
      "1793960115486236481",
      "1793960115486236494",
      "1793960115486236489",
      "1793960115486236490"
    ],
    "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/1793997239974447410",
      "all": "for id in 1793997239974447410 1793960115486236481 1793960115486236494 1793960115486236489 1793960115486236490; do\n  curl -s -H \"Authorization: Bearer $SYNORB_KEY\" \"https://api.synorb.com/manifests/by-id/$id\"\ndone"
    }
  },
  "links": {
    "html": "https://hangingcontext.com/stories/2026-11-07/agents-that-improve-themselves/",
    "json": "https://hangingcontext.com/stories/2026-11-07/agents-that-improve-themselves/index.json",
    "markdown": "https://hangingcontext.com/stories/2026-11-07/agents-that-improve-themselves/index.md",
    "newsroom": "https://hangingcontext.com/n/artificial-intelligence/",
    "edition": "https://hangingcontext.com/editions/2026-11-07T00/"
  },
  "related": [
    {
      "url": "https://hangingcontext.com/stories/2026-11-07/being-cited-by-agents/",
      "headline": "Brands now compete to be the answer the agent cites",
      "slug": "being-cited-by-agents"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-06/voice-ai-money-story-arrives/",
      "headline": "Voice AI's money story arrives: record revenue and a real-time model",
      "slug": "voice-ai-money-story-arrives"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-06/drug-discovery-data-wins-benchmarks/",
      "headline": "In drug discovery, the data wins the benchmark",
      "slug": "drug-discovery-data-wins-benchmarks"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-06/customer-service-ai-consolidation-week/",
      "headline": "Five9 buys Level AI and Sierra buys Anon as customer-service AI consolidates",
      "slug": "customer-service-ai-consolidation-week"
    }
  ],
  "sponsor": {
    "label": "From Synorb",
    "title": "Run the cited Manifests yourself",
    "text": "Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. 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.",
    "cta": {
      "label": "Get a key",
      "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/2026-11-07/agents-that-improve-themselves/"
  }
}