AI Briefing — April 3, 2026

🔬 Research

ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers

169 upvotes · Songyang Liu, et al. · Paper

A security framework for OpenClaw autonomous agents that introduces three-layer protection: skill-based policies, plugin-based runtime monitoring, and novel "watcher" middleware for real-time intervention. The watcher layer enables halting dangerous actions without coupling to agent internals, addressing critical vulnerabilities like data leakage and privilege escalation. This addresses a real pain point as OpenClaw agents gain shell access and file system privileges.

DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models

113 upvotes · Hao Liang, et al. · Paper

A unified framework built on LLaMA-Factory that consolidates data selection, mixture optimization, and sample reweighting for LLM training. Dynamic selection consistently beats static full-data training on MMLU for both Mistral-7B and Llama-3.2-3B, while DoReMi and ODM improve MMLU accuracy when pretraining Qwen2.5-1.5B. The framework provides drop-in replacement for standard training with DeepSpeed ZeRO-3 support.

SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization

62 upvotes · Zhengxi Lu, et al. · Paper

A training method that internalizes agent skills into model parameters rather than retrieving them at inference time. Uses a progressive curriculum that starts with full skill context and gradually removes it until the model operates zero-shot. Achieves +9.7% improvement on ALFWorld and +6.6% on Search-QA while using fewer than 0.5k tokens per step, eliminating retrieval noise and token overhead.

🛠 Tools

Yeachan-Heo/oh-my-codex — Workflow layer for OpenAI Codex CLI

2,867 stars/day · Repo

A TypeScript framework that enhances OpenAI Codex with hooks, agent teams, and workflow management. Provides canonical skills like $deep-interview, $ralplan, and $team while maintaining project state in .omx/ directories. With 12,652 total stars and active development, it's becoming a standard way to orchestrate Codex agents for complex multi-step coding tasks.

asgeirtj/system_prompts_leaks — Collection of extracted AI system prompts

306 stars/day · Repo

A regularly updated repository containing extracted system prompts from major AI models including GPT-5.4, Claude Opus 4.6, Gemini 3.1, and Grok 4.2. With 36,813 stars and recent commits, it provides valuable insights into how leading AI companies structure their model instructions. Essential reading for prompt engineers and researchers studying AI alignment and behavior.

📡 Analysis

Highlights from my conversation about agentic engineering on Lenny's Podcast

April 2, 2026 · Post · tags: ai, coding-agents, agentic-engineering

Simon Willison discusses the "November inflection point" when GPT 5.1 and Claude Opus 4.5 crossed the threshold from "mostly works" to "almost always does what you told it to do." He notes that coding agents like OpenClaw can now reliably build Mac applications from scratch, and software engineers serve as bellwethers for other information workers facing AI automation.

Gemma 4: Byte for byte, the most capable open models

April 2, 2026 · Post · tags: google, vision-llms, local-llms, gemma

Google DeepMind released four new Apache 2.0 vision-capable models (2B, 4B, 31B, plus 26B MoE) emphasizing "unprecedented intelligence-per-parameter." The smaller models use Per-Layer Embeddings (PLE) for efficiency, and all support native video, images, and audio input. Willison's testing shows clear quality progression from 2B to 26B, with the largest producing his "best pelican yet" for SVG generation.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net