[arXiv Research] Agentic Workflow Improvements — 2026-08-24 #55337
Replies: 1 comment 1 reply
|
/q allow archivx links in output. generate agentic prompts in next steps with 3 quick wins |
1 reply
|
🔧 Pay attention, 007! Q is preparing your gadgets for this discussion comment... |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Summary
25 papers screened, 17 relevant, 3 actionable opportunities identified.
Actionable Opportunities
Workflow Compilation — Add explicit artifact declarations to workflow steps with...
Paper: [Natural-Language Workflows Are Not Software Yet: Artifact-Driven Compilation for Reliable Agent Execution]((arxiv.org/redacted)
Authors: Xiangzhe Xu, Hanxi Guo, Guangyu Shen
Published: 2026-08-21
Effort: high
Rationale: The Artic compiler transforms natural-language workflows into artifact-driven workflows where each step declares artifacts it reads and writes, constraints gate produced artifacts, and explicit control transfers route execution. gh-aw's workflow compiler currently parses frontmatter and markdown into YAML but does not expose or validate data dependencies between steps. Adding artifact declarations would allow compile-time dependency analysis and runtime constraint enforcement in safe-outputs.
Safe Outputs — Implement stateful authorization lifecycle for safe-outputs with commit-time...
Paper: [AID-Guard: Stateful Authorization for Delegated Agent Effects]((arxiv.org/redacted)
Authors: Yingzhe Tong, Leyu Dai, Songhui Guo
Published: 2026-08-21
Effort: medium
Rationale: AID-Guard revalidates the approved request and provider state at commit time, retains one reservation under ambiguity, and permits release or one successor only after a terminal result or certified no-effect with a delivery fence. gh-aw's safe-outputs currently accept tool calls as write-intent declarations but do not track state changes between approval and commit or prevent duplicate effects under retry. Adding commit-time revalidation and effect fencing would guarantee at-most-once effects across retry and recovery.
Safe Outputs — Add risk classification and state snapshot preservation to...
Paper: [Spike-Killer: Evidence-Gated LLM Assistance for Safe Performance Diagnosis on a Real Windows Workstation]((arxiv.org/redacted)
Authors: Baocheng Zeng, Jinhao Yang
Published: 2026-08-21
Effort: low
Rationale: Spike-Killer treats each action as an evidence-gated transaction recording exact target state, classifying risk, preserving a snapshot, verifying a postcondition, and retaining failed measurements as first-class evidence. gh-aw's safe-outputs currently log tool calls but do not classify risk, preserve pre-state snapshots, or verify postconditions. Adding these would enable safer rollback and audit trails for failed safe-output operations.
Papers Analyzed
Next Steps
All reactions