[arXiv Research] Agentic Workflow Improvements — 2026-09-30

[github-actions[bot]]

Screened 25 new arXiv papers by title and abstract. I did not run the PDF investigation, screener, ranker or extractor sub-agents, and the ledger and dedup cache were not updated (see the last section). Three papers looked relevant.

Opportunities

1. Do LLM Agents Execute the Plans They Declare? (2609.38108)

   • Agents often fail to follow the plan they declare. Compare the declared plan with the executed steps in run logs, and audit that gap. Consider routing to pattern-specific executors.
   • Area: agent orchestration / observability.

2. Merged, Not Measured: Performance Issues Fixed by Coding Agents (2609.37985)

   • In the paper's sample, 57% of closed agent performance fixes were merged. 61% of rejections gave no stated reason. Only 6 of 23 re-executed rejected claims held.
   • Require safe-output PRs and issues that make performance claims to include re-runnable benchmark evidence.
   • Area: safe outputs.

3. AgentBug-Smith: Reproducing Real-World Harness Bugs (2609.37864)

   • The paper reproduces harness bugs from open-source agentic systems automatically. Use similar reproduction to build regression tests for engine and harness failures in compiled workflows.
   • Area: testing / engines.

Ledger status

The sandbox denied bash python execution, so paper-ledger.md, paper-index.json and seen-paper-ids.json were not updated. The next run may re-process these 25 papers.

Generated by 🔬 arXiv Paper Researcher: GitHub Agentic Workflows · copilot · large · 33 AIC · ⌖ 13.4 AIC · ⊞ 10.4K · ◷

• expires on Oct 14, 2026, 12:45 AM UTC-08:00