AI Daily — June 16, 2026
2026-06-16
code.claude.com
Claude Code Week 24: Recursive Sub-agents, Mid-session Directory Changes, and Fallback Models
Claude Code now supports sub-agents spawning their own sub-agents up to five levels deep, enabling more complex background agentic chains. The new /cd command lets users relocate a session's working directory mid-conversation without invalidating the prompt cache, and fallbackModel accepts an ordered list of up to three fallback models for resilience. A --safe-mode flag disables all customizations for clean troubleshooting.
code.claude.com
Claude Code Week 23: Auto Mode on Third-Party Providers and Safer Automatic Edits
Auto mode — which replaces interactive permission prompts with background safety checks — is now available on AWS Bedrock, Google Vertex, and Azure Foundry for Opus 4.7 and Opus 4.8. In acceptEdits mode, the agent now prompts before writing executable files, reducing the risk of unintended code execution during automated edit sessions.
jack-clark.net
Import AI 461: Researchers Launch Safety Startup 'Sequent' Citing Alignment Off Track
Researchers from the UK AI Security Institute have spun out a new safety-focused startup called Sequent, explicitly stating that alignment research is not on track to keep pace with frontier model development. The organization aims to fund under-resourced research bets in alignment and safety. The issue also covers FrontierCode, a coding benchmark, and experiments using LLMs as synthetic research interns.
rss.arxiv.org
μ₀: Embodiment-Agnostic 3D Interaction-Trace World Model for Robot Learning
μ₀ is a world model that predicts smooth 3D trajectories for salient interaction points (objects, hands, tools, contact regions) rather than dense pixels or embodiment-specific action labels, making it scalable across diverse robot morphologies. The accompanying TraceExtract system automatically derives these traces from raw video, enabling training on large heterogeneous video datasets without embodiment-specific annotation. This compact motion interface bridges the gap between general video pretraining and deployable robot policy learning.