[arXiv Research] Agentic Workflow Improvements — 2026-10-16

[github-actions[bot]]

Summary

25 papers screened, 22 relevant, 3 opportunities identified.

---

Actionable Opportunities

Runtime orchestration — Opt-in goal-drift steering for long-running workflows

Paper: Learning to Steer LLM Agents and Multi-Agent Systems
Authors: Leyang Shen, Yaxin Gong, Yang Zhang
Published: 2026-10-15
Effort: high
Rationale: Long tool-using workflows can spend substantial tokens and time pursuing locally plausible actions that no longer advance the original goal. Add an opt-in workflow steering policy where a low-cost engine emits a typed drift priority and only priorities above a configured threshold invoke a stronger steering engine with bounded corrective guidance.

---

Evaluator reliability — Add shortcut audits to operational-value validation

Paper: Shortcut Sensitivity in Agent-Safety Evaluations
Authors: Hasan Demirkiran, Jens Ernstberger
Published: 2026-10-15
Effort: medium
Rationale: An enabled evaluator may reward superficial workflow properties or repository/domain cues, making apparent improvements fail to generalize while still satisfying ordinary fixtures. Extend the operational-value designer workflow with prospective diagnostics for shallow-cue predictability, domain-held-out performance, and paired decision stability while preserving the evaluator/workflow contract and all prior results.

---

Sub-agent orchestration — Dependency-aware hierarchical sub-agent plans

Paper: HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration
Authors: Yanchen Deng, Jianguo Huang, Yao Long Teng
Published: 2026-10-15
Effort: high
Rationale: gh-aw can delegate work, but prompts and static workflow structure do not natively represent nested tasks, dependencies, readiness, or upward result synthesis. Add an opt-in hierarchical task-plan abstraction whose nodes decompose only when prerequisites are ready, whose leaves dispatch sub-agents, and whose parents complete under explicit AND semantics.

---

Papers Analyzed

Paper	Published	Relevant	Area
LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation	2026-10-15	Yes	evaluation semantics
FDE-Gym: The Last Mile of AI Delivery	2026-10-15	Yes	agent evaluation
MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering	2026-10-15	Yes	trajectory evaluation
3DCodeVerse: Exploring 3D Worlds as Code	2026-10-15	Yes	artifact verification
KnowGPO: Knowledge-Grouped Policy Optimization for Search Agents	2026-10-15	Yes	token optimization
Learning Self-Orchestration in Decentralized Agent Swarms	2026-10-15	Yes	sub-agent orchestration
A team of Gemini 4 Argon agents beats Minecraft	2026-10-15	Yes	durable multi-agent execution
Learning to Steer LLM Agents and Multi-Agent Systems	2026-10-15	Yes	runtime steering
CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty	2026-10-15	Yes	trustworthy completion
Improving Environment-Agnostic Coding Agents: The Role of Scaffolding	2026-10-15	Yes	prompts and continuation
Learning Rational Strategies for Temporal Objectives in Multi-Agent Systems	2026-10-15	No	—
LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents	2026-10-15	Yes	engine routing and oversight
Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries	2026-10-15	Yes	agent memory
X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning	2026-10-15	No	—
Disagreement-Guided Weak-to-Strong Routing	2026-10-15	Yes	engine routing and token cost
Language Is a Security Surface for Agent Guards	2026-10-15	Yes	safe-output security
Shortcut Sensitivity in Agent-Safety Evaluations	2026-10-15	Yes	security evaluation
OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents	2026-10-15	Yes	engine routing
Architecting for Hardware Diversity: Meta’s End-to-End Co-Design for High-Performance AI Training on AMD GPUs	2026-10-15	Yes	engine portability
FutureLift: Lifting Trajectory Proposals into Bidirectional Future Graphs for Motion Forecasting	2026-10-15	No	—
Beyond Downstream QA: Rethinking How We Evaluate Agent Memory	2026-10-15	Yes	memory evaluation
Decentralized LLM Workflows via Sequential Auctions	2026-10-15	Yes	sub-agent routing
HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration	2026-10-15	Yes	sub-agent orchestration
AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning	2026-10-15	Yes	token optimization
Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine	2026-10-15	Yes	retry evaluation

---

Next Steps

• Investigate: opt-in goal-drift steering for long-running workflows (effort: high)
• Investigate: shortcut audits for operational-value validation (effort: medium)
• Investigate: dependency-aware hierarchical sub-agent plans (effort: high)

Quick-Win Agentic Prompts

@copilot Implement: Add an opt-in workflow steering policy where a low-cost engine emits a typed drift priority and only priorities above a configured threshold invoke a stronger steering engine with bounded corrective guidance. in gh-aw’s runtime orchestration component. Rationale: Long tool-using workflows can spend substantial tokens and time pursuing locally plausible actions that no longer advance the original goal. Add an opt-in workflow steering policy where a low-cost engine emits a typed drift priority and only priorities above a configured threshold invoke a stronger steering engine with bounded corrective guidance. Source: Learning to Steer LLM Agents and Multi-Agent Systems.

@copilot Implement: Extend the operational-value designer workflow with prospective diagnostics for shallow-cue predictability, domain-held-out performance, and paired decision stability while preserving the evaluator/workflow contract and all prior results. in gh-aw’s evaluator reliability component. Rationale: An enabled evaluator may reward superficial workflow properties or repository/domain cues, making apparent improvements fail to generalize while still satisfying ordinary fixtures. Extend the operational-value designer workflow with prospective diagnostics for shallow-cue predictability, domain-held-out performance, and paired decision stability while preserving the evaluator/workflow contract and all prior results. Source: Shortcut Sensitivity in Agent-Safety Evaluations.

@copilot Implement: Add an opt-in hierarchical task-plan abstraction whose nodes decompose only when prerequisites are ready, whose leaves dispatch sub-agents, and whose parents complete under explicit AND semantics. in gh-aw’s sub-agent orchestration component. Rationale: gh-aw can delegate work, but prompts and static workflow structure do not natively represent nested tasks, dependencies, readiness, or upward result synthesis. Add an opt-in hierarchical task-plan abstraction whose nodes decompose only when prerequisites are ready, whose leaves dispatch sub-agents, and whose parents complete under explicit AND semantics. Source: HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration.

References: Workflow run §38874996797

Generated by 🔬 arXiv Paper Researcher: GitHub Agentic Workflows · codex · gpt56 · 193.3 AIC · ⌖ 13 AIC · ⊞ 18.1K · ◷

• expires on Oct 30, 2026, 12:45 AM UTC-08:00