AI Roundup: Cursor 3 Goes Multi-Agent, Safetensors Joins PyTorch Foundation, and Agents That Actually Learn

Three AI stories surfaced today across Hacker News, Simon Willison's blog, and the Hugging Face blog. Here's what's worth your time.

TL;DR: Cursor 3 ships an Agents Window for parallel development. Safetensors moves to the PyTorch Foundation. IBM's ALTK-Evolve gives agents real working memory instead of re-reading logs.

Cursor 3 Ships an Agents Window

Cursor released a major update today. The headline feature is the new Agents Window, which lets you run multiple AI agents in parallel across different repos, worktrees, cloud environments, and remote SSH machines.

This is a shift from the single-agent model that dominates most AI coding tools right now. Instead of waiting for one agent to finish before starting another, you can spin up several at once and have them work simultaneously on different parts of your codebase.

The new interface also includes Design Mode for targeting UI elements directly in the browser, Agent Tabs for viewing multiple chats side-by-side, and improved Bugbot with learned rules. Bugbot now reads PR feedback and turns it into reusable rules that improve future code reviews. It also gains MCP support, meaning it can tap into external tools during reviews.

If you're a Cursor user, the upgrade is available now. Type Cmd+Shift+P -> Agents Window to try it.

Safetensors Joins the PyTorch Foundation

Safetensors, the model weight format that became the de facto standard for sharing open models, is now under the PyTorch Foundation. The move puts the project under Linux Foundation governance, separate from any single company.

The format launched as a Hugging Face project years ago, built specifically to replace pickle-based model files that could execute arbitrary code. The core design is simple: a JSON header describing tensor metadata, followed by raw weight data. Zero-copy loading, lazy loading, no serialization overhead.

Today it's used by tens of thousands of models across the Hugging Face Hub. Every Llama variant, every Stable Diffusion checkpoint, every embedding model is probably sitting on disk as a Safetensors file right now.

The interesting part: nothing changes for users. The APIs, the format, the Hub integration all stay the same. But contributors now have a formal path to become maintainers, and the project has neutral governance instead of being tied to one organization. Luc and Daniel, the core maintainers, remain on the technical steering committee.

The roadmap includes device-aware loading (directly onto CUDA, ROCm), tensor and pipeline parallel loading, and formal support for FP8 and block-quantized formats like GPTQ and AWQ.

ALTK-Evolve: Teaching Agents to Actually Learn

IBM Research published a paper on a problem that has plagued AI agents since day one. Most agents re-read their own logs instead of learning from them. Every session is a reset. Feed yesterday's conversation into today's prompt and you get noise, not wisdom.

ALTK-Evolve is a memory system that converts agent trajectories into reusable guidelines. Instead of stuffing context windows with past conversations, it extracts principles from past runs and injects only the relevant ones at the moment of action.

The architecture has two flows. Downward: capture full agent traces (user input, thoughts, tool calls, results) and mine them for structural patterns. Upward: consolidate, score, and merge duplicates into a library of guidelines. When a new task comes in, only the most relevant guidelines get retrieved.

The results on AppWorld benchmarks show the harder the task, the bigger the lift. Hard tasks saw a 14.2% absolute improvement in goal completion. The agent isn't memorizing; it's generalizing.

For developers building on top of agent frameworks, this is relevant. ALTK-Evolve has integrations with Claude Code, Codex, and IBM Bob in lite mode, plus low-code tracing for ReAct agents and an MCP integration for CUGA.

The 95% figure from an MIT study cited in the paper is stark. Most AI pilots fail because agents don't adapt on the job. This approach is one attempt to fix that.

Sources: Cursor Changelog, Hugging Face Blog, IBM Research, Simon Willison