[arXiv Research] Agentic Workflow Improvements — 2026-08-23

[github-actions[bot]]

Summary

25 papers screened, 14 relevant, 3 high-priority opportunities identified.

---

Actionable Opportunities

Token Optimization — Optimal Skill Selection with Budget Constraints

Paper: [Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees]((arxiv.org/redacted)
Authors: Yu Chen, Ruishuo Chen, Xun Wang
Published: 2026-08-20
Effort: medium
Rationale: The paper presents a provably optimal algorithm for selecting skills under hard token budgets by modeling benefit as monotone submodular and penalizing redundancy. This directly improves gh-aw's skill fusion and lazy loading system, which currently lacks formal cost-benefit optimization.

---

Workflow Compilation — MCP-Aware Training Data Pipeline

Paper: [MidTool: Mid-training Data Synthesis for Agentic Tool Use]((arxiv.org/redacted)
Authors: Fengqing Jiang, Yite Wang, Boyi Liu
Published: 2026-08-20
Effort: high
Rationale: MidTool demonstrates that synthesizing training data from real tool APIs, MCP skills, and document-grounded workflows teaches models tool affordance recognition and workflow composition. This could produce gh-aw-specialized models that better understand workflow markdown syntax, safe-outputs constraints, and multi-agent patterns.

---

Prompt Engineering — Instruction-First Documentation Structure

Paper: [From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation]((arxiv.org/redacted)
Authors: Zhijun Gao, Jing Chen
Published: 2026-08-20
Effort: low
Rationale: The study shows agents interact with instruction files and working notes 60.5% of the time versus 10.6% for traditional docs. gh-aw already uses SKILL.md as instruction files, but could formalize compact command playbooks, example workflows, and decision trees at the top, moving detailed API reference lower.

---

Papers Analyzed

Paper	Published	Relevant	Area
[An Agentic Approach for Active Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Prediction]((arxiv.org/redacted)	2026-08-20	No	—
[Inducing Task Models from Computer-Use Traces]((arxiv.org/redacted)	2026-08-20	Yes	multi-agent
[AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement]((arxiv.org/redacted)	2026-08-20	Yes	other
[MidTool: Mid-training Data Synthesis for Agentic Tool Use]((arxiv.org/redacted)	2026-08-20	Yes	workflow-compilation
[Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents]((arxiv.org/redacted)	2026-08-20	Yes	multi-agent
[Growth Without Us: Machine Consumers, Corporate Circularity, and the Decoupling of GDP from Humanity after AGI]((arxiv.org/redacted)	2026-08-20	No	—
[The Third Restructuring of Software Form: From the Three-Tier Architecture to Storage, Models, and Agents]((arxiv.org/redacted)	2026-08-20	Yes	other
[From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation]((arxiv.org/redacted)	2026-08-20	Yes	prompt-engineering
[Task-CoEvolve: Efficient Harness Optimization via Adaptive Validation Task Selection]((arxiv.org/redacted)	2026-08-20	Yes	other
[Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design]((arxiv.org/redacted)	2026-08-20	Yes	multi-agent
[A three-dimensional typology of agency for advanced AI systems]((arxiv.org/redacted)	2026-08-20	No	—
[Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees]((arxiv.org/redacted)	2026-08-20	Yes	token-optimization
[ReguSim: Evaluating LLM Agent Rule Grounding in Financial Compliance]((arxiv.org/redacted)	2026-08-20	Yes	security
[G-MARK: Grounded Multi-Agent Reasoning for Cooperative Driving via Knowledge Graphs]((arxiv.org/redacted)	2026-08-20	No	—
[Bringing analytic rigor to agentic AI for science: The Brain Researcher platform for neuroimaging data analysis]((arxiv.org/redacted)	2026-08-20	No	—
[MaliciousSkillBench: A Comprehensive Benchmark for Malicious Agent Skill Detection]((arxiv.org/redacted)	2026-08-20	Yes	security
[TESTNAV: Pareto-Guided Search for Compositional Robustness Testing]((arxiv.org/redacted)	2026-08-20	No	—
[EnvHarness: Awakening Static Worlds for Agent Learning]((arxiv.org/redacted)	2026-08-20	Yes	other
[A knowledge-guided agentic framework for mitigating patient-context ambiguity in health queries]((arxiv.org/redacted)	2026-08-20	No	—
[PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents]((arxiv.org/redacted)	2026-08-20	Yes	security
[Inadvertent Context Leakage in Language Models]((arxiv.org/redacted)	2026-08-20	Yes	security
[Repo0: Design-Driven Zero-to-All Code Generation]((arxiv.org/redacted)	2026-08-20	No	—
[SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning]((arxiv.org/redacted)	2026-08-20	No	—
[Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning]((arxiv.org/redacted)	2026-08-20	No	—
[MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents]((arxiv.org/redacted)	2026-08-20	No	—

---

Next Steps

• Investigate: Optimal skill selection optimization (effort: medium)
• Investigate: MCP-aware training pipeline for gh-aw-specialized models (effort: high)
• Investigate: Instruction-first SKILL.md restructuring (effort: low)

Generated by 🔬 arXiv Paper Researcher: GitHub Agentic Workflows · sonnet45 · 73 AIC · ⊞ 3.9K · ◷

• expires on Sep 6, 2026, 12:42 AM UTC-08:00

[pelikhan]

/q add subagents to do preliminary investigation. Use agentic engine that can read PDFs or use MCP to open PDF.
Allow archiv URLS in report

[github-actions[bot]]

🔧 Pay attention, 007! Q is preparing your gadgets for this discussion comment...