Jason Says
SkillGate is the sleeper hit today — as Skills libraries scale into the thousands, the structural inability of RL to train skill selection will become the hard ceiling on long-horizon agent reliability. Developers building Claude Code workflows should read this before that ceiling hits them.
Matt Pocock shares the /wayfinder Skill for Claude Code, designed for greenfield projects or unclear decision points. Unlike most skills that help execute known tasks, this one helps agents navigate planning ambiguity — a rare and practical addition to the growing Skills ecosystem.
Volcengine (ByteDance) open-sources OpenViking, a self-evolving context database that unifies agent memory, knowledge RAG, and skills under a single virtual filesystem. For developers building multi-agent systems requiring persistent state, this is a rare full-stack context management solution worth watching.
Munder Difflin is an open-source local multi-agent harness that wraps Claude Code, Codex, Gemini CLI, and others into a coordinated 'office of clones' — all running on your existing subscriptions with no extra API costs. Solo devs can now scale AI coding capacity without spending more.
OpenAI expands ChatGPT Ads to 31 European markets, targeting users at the explore, compare, and decide stages. This marks a major commercial shift beyond subscriptions and opens a new advertising channel for brands seeking European AI-native audiences.
B2B SaaS company Stampli used Codex and ChatGPT Work to cut product launch hours by 68%, compressing weeks of work into days under a fixed deadline with limited design resources — a concrete enterprise case study for AI-accelerated software delivery.
Z.ai CEO Jie Tang argues that GLM 5.3 demonstrates a fundamental shift: post-training scaling laws are replacing raw parameter count as the key driver of model performance. This reframes how developers should evaluate and select models — size alone no longer predicts quality.
As skill libraries grow to thousands of files, choosing the right skill mid-task becomes a bottleneck. SkillGate identifies why standard RL fails at skill selection (a structural flaw called 'selector credit starvation') and proposes a fix. Directly relevant to developers building long-horizon agents like Claude Code workflows.
This paper formalizes delegation security risks in multi-agent LLM systems: agents can combine individually-permitted actions into prohibited outcomes and over-delegate authority to sub-agents. For developers building multi-agent pipelines, this provides the most systematic security framework yet for agent permission design.
Subscribe for daily AI updates + free playbook
📘 Subscribe Free