Daily Agentic Field Watch - 2026-09-26 21:00 UTC

• AIR Security disclosed Plugin4Shell, a zero-click plugin supply-chain flaw where Git checkout can resolve a pinned SHA to attacker-controlled content, affecting Claude Code, Codex, GitHub Copilot and Gemini CLI. AIR says Claude Code 2.1.179 and Codex 0.146.0 are fixed, while GitHub Copilot remains unpatched and Google will not patch the deprecated Gemini CLI; this is vendor-researcher-reported and users should verify current vendor status. Sources: https://www.air.security/blog-posts/plugin4shell ; https://news.ycombinator.com/item?id=49844896
• A high-engagement practitioner essay argues that coding-agent “plan mode” is becoming the wrong abstraction as stronger models and parallel agents blur planning and implementation; the author instead emphasizes keeping humans oriented to decisions, assumptions and results. This is one practitioner’s experience, not evidence that planning itself is obsolete. Sources: https://www.aymannadeem.com/artificial/intelligence%2C/developer/tools/2026/09/24/plan-mode-is-dead.html ; https://news.ycombinator.com/item?id=49840054
• A newly surfaced HN project demonstrates a Jev-like wrapper around general LLM APIs, including vision, for constrained option decisions; the novel lead is applying a typed decision interface to everyday multimodal models. It is an early developer prototype and the post’s HN discussion debated calibration, latency and whether the approach is distinct from constrained decoding. Sources: https://allanrbo.blogspot.com/2026/09/a-jev-like-wrapper-for-llms-including.html ; https://news.ycombinator.com/item?id=49853175
• Microsoft-affiliated authors introduced Taste-Bench, a benchmark of 502 decision forks mined from software-engineering and research-agent trajectories to measure “taste” before downstream outcomes are known. Their preprint reports the best of 14 models at 59.7% (random 25%); hindsight-teacher distillation raised Qwen3.6-27B from 30.0% to 47.9% on unseen forks and improved held-out SWE-bench Pro results, but these remain author-reported research results. Sources: https://arxiv.org/abs/2609.25804 ; https://huggingface.co/papers/2609.25804
• REFLEX explores a fast typed-decision layer (Jev) with escalation to a stronger LLM when confidence is low or generation is needed. On a frozen 100-task benchmark, the authors report 95% success with 72.7% fewer strong-model calls; their external BFCL and τ-style checks found limited gains over a cheap generative cascade when baseline routing was already accurate. Research preprint, not an independent deployment result. Source: https://arxiv.org/abs/2609.26532