{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2609.09153v1",
  "slug": "2609-09153v1-0ee5zbi",
  "url": "https://feed7.dev/p/2609-09153v1-0ee5zbi",
  "title": "Procedural Graphs: Self-Evolving Execution Structures for LLM Agents",
  "why_included": "Procedural Graphs externalize an agent’s workflow, guide each action from the relevant subgraph, and refine the structure from failed and working runs. This targets drift and repeated tool mistakes on long tasks.",
  "summary": "A **Procedural Graph** represents what to do as linked procedure-relation-procedure triples. At each decision, the system locates the active node and converts its surrounding subgraph into guidance for the solver’s next action.",
  "practical_implication": "For long-running coding agents, this suggests making workflow state explicit instead of relying only on an expanding transcript. The refinement loop can compare failed and working trajectories, then update the graph while preserving changes that help held-out validation.",
  "agent_context": "A **Procedural Graph** represents what to do as linked procedure-relation-procedure triples. At each decision, the system locates the active node and converts its surrounding subgraph into guidance for the solver’s next action.\n\nFor long-running coding agents, this suggests making workflow state explicit instead of relying only on an expanding transcript. The refinement loop can compare failed and working trajectories, then update the graph while preserving changes that help held-out validation.\n\nThe material reports consistent gains over memory-based baselines across multiple datasets, task types, and models, but gives no result sizes here. The approach also adds a guidance model, a refiner, and validation machinery to the harness.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2609.09153v1",
    "published_at": "2026-09-08T17:59:41.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "agent",
  "domains": [
    "coding"
  ],
  "topics": [
    "harness-engineering",
    "agent-memory",
    "agent-reliability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The material reports consistent gains over memory-based baselines across multiple datasets, task types, and models, but gives no result sizes here. The approach also adds a guidance model, a refiner, and validation machinery to the harness."
  ],
  "connected_context": {
    "meaning": "This turns durable agent state from a passive record into an executable, locally retrieved workflow that can be revised from outcomes. It gives the prior long-running harness patterns a concrete adaptive structure, while raising the bar for curation: graph updates need held-out validation so accumulated experience does not encode regressions.",
    "corpus_size": 713,
    "generated_at": "2026-09-09T10:12:53.401Z",
    "connections": [
      {
        "title": "Effective harnesses for long-running agents",
        "source_name": "Anthropic",
        "source_url": "https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents",
        "feed7_url": "https://feed7.dev/p/effective-harnesses-for-long-running-agents-0xzfs05",
        "reason": "Procedural Graphs extend Anthropic’s explicit feature and progress files into linked action guidance that can evolve from trajectories."
      },
      {
        "title": "Chained Recursive Language Models for Multi-Iteration Reasoning",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.05124v1",
        "feed7_url": "https://feed7.dev/p/2608-05124v1-05haobv",
        "reason": "Both externalize state for long-horizon work, but Chained RLM carries summaries and artifacts between fresh contexts while Procedural Graphs retrieve and refine an executable neighborhood."
      },
      {
        "title": "The Era of Compound Engineering — Kieran Klaassen, Every/Cora",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk",
        "feed7_url": "https://feed7.dev/p/the-era-of-compound-engineering-kieran-klaassen-every-cora-053chml",
        "reason": "The graph operationalizes compound engineering’s reusable corrections, while held-out validation supplies a safeguard against memory that merely accumulates."
      },
      {
        "title": "A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.02553v1",
        "feed7_url": "https://feed7.dev/p/2608-02553v1-12y8joy",
        "reason": "It supplies a concrete adaptive-state architecture for several reliability needs that the capability-gap taxonomy identifies only conceptually."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-08T17:59:41.000Z",
  "modified_at": "2026-09-08T17:59:41.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2609-09153v1-0ee5zbi",
    "json": "https://feed7.dev/p/2609-09153v1-0ee5zbi.json",
    "markdown": "https://feed7.dev/p/2609-09153v1-0ee5zbi.md"
  }
}