{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2608.06346v1",
  "slug": "2608-06346v1-125r9ui",
  "url": "https://feed7.dev/p/2608-06346v1-125r9ui",
  "title": "TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories",
  "why_included": "TrajDebug tracks whether errors persist, resolve, or cause terminal failure, offering a sharper way to debug long coding-agent runs than flagging every local mistake.",
  "summary": "**TrajDebug** compresses long histories at multiple granularities, identifies errors from evidence, then tracks resolution status and terminal impact. **TrajErrBench** contains **486 manually annotated failed trajectories** from Tau2Bench and SWE-Bench Pro.",
  "practical_implication": "For coding-agent observability, preserve enough trajectory context to distinguish a harmless local mistake from the earliest unresolved error that caused the final failure. That attribution can make remediation more targeted.",
  "agent_context": "**TrajDebug** compresses long histories at multiple granularities, identifies errors from evidence, then tracks resolution status and terminal impact. **TrajErrBench** contains **486 manually annotated failed trajectories** from Tau2Bench and SWE-Bench Pro.\n\nFor coding-agent observability, preserve enough trajectory context to distinguish a harmless local mistake from the earliest unresolved error that caused the final failure. That attribution can make remediation more targeted.\n\nThe abstract claims the best overall performance and actionable downstream feedback but gives no scores or baseline details. Code and data are promised for release, so reproducibility cannot be assessed from this material.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2608.06346v1",
    "published_at": "2026-08-06T17:51:20.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [
    "coding"
  ],
  "topics": [
    "agent-evals",
    "agent-reliability",
    "observability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The abstract claims the best overall performance and actionable downstream feedback but gives no scores or baseline details. Code and data are promised for release, so reproducibility cannot be assessed from this material."
  ],
  "connected_context": {
    "meaning": "This turns long-trajectory inspection into lifecycle attribution: identify errors from evidence, track whether they were resolved, and isolate the earliest unresolved mistake linked to terminal failure. It operationalizes prior calls for queryable trajectories and flexible analysis, while going beyond pass/fail verification toward targeted remediation. The absent scores, baseline details, code, and data keep the claimed advantage provisional.",
    "corpus_size": 390,
    "generated_at": "2026-08-08T10:06:15.427Z",
    "connections": [
      {
        "title": "The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=q2JrUKBMf0w",
        "feed7_url": "https://feed7.dev/p/the-future-of-evals-from-llm-as-a-judge-to-agent-as-a-judge-aparna-dhina-1fu560o",
        "reason": "TrajDebug supplies a concrete agent-analysis task for failures that fixed rubrics miss: follow errors across a variable trajectory and distinguish resolved mistakes from terminally consequential ones."
      },
      {
        "title": "Everything Is a Rollout — Alex Shaw + Ryan Marten, Terminal-Bench, Harbor, Laude Institute",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=jRCpXUjz4CI",
        "feed7_url": "https://feed7.dev/p/everything-is-a-rollout-alex-shaw-ryan-marten-terminal-bench-harbor-laud-0iz4rgx",
        "reason": "Harbor’s reproducible rollout loop provides the trajectories and verified outcomes that TrajDebug needs; TrajDebug adds causal error attribution to their inspection."
      },
      {
        "title": "Rethinking Environments for Long-Horizon Work — Rayan Garg, Theta Software",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=2aS7aKoXn64",
        "feed7_url": "https://feed7.dev/p/rethinking-environments-for-long-horizon-work-rayan-garg-theta-software-11r7wbx",
        "reason": "Multi-granularity history compression is an implementation response to the requirement that long traces remain queryable, while terminal-impact tracking preserves the link to final-state evaluation."
      },
      {
        "title": "Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code Models",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.12962v1",
        "feed7_url": "https://feed7.dev/p/2607-12962v1-0q4i26c",
        "reason": "The placebo-controlled repair result warns that feedback labeled as an error may not be causally informative; TrajDebug’s evidence-based attribution could help distinguish actionable failure content, though its remediation benefit remains unquantified."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-08-06T17:51:20.000Z",
  "modified_at": "2026-08-06T17:51:20.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2608-06346v1-125r9ui",
    "json": "https://feed7.dev/p/2608-06346v1-125r9ui.json",
    "markdown": "https://feed7.dev/p/2608-06346v1-125r9ui.md"
  }
}