OpenHands: separating the software agent from the sandbox it works in
A new healthy-local-backend fallback makes the project’s larger argument concrete: the thing that helps a person steer an agent must stay legible even when the agent is running somewhere else.
More From This Edition
All editions ›The Harness Was Taking the Test Too
Builder Work: A Sandbox Is Reusable Only After It Forgets
API Discovery: Make the Model Own Its Web Search
Recent Editions
Full history ›Hugging Face Links 3D Dataset Previews to Their Mesh Files
A newly merged dataset-viewer change pairs a flat picture with a retrievable 3D asset. It connects a quick look at a dataset row to the geometry behind it—while wider “3D” labels remain a separate capability.
Unsloth Connects Existing Fine-Tuning Scripts to Multiple GPUs
Supported training scripts can now divide model work across graphics cards through a launch setting. A later change reconnects training to generation and export—with a different memory requirement at each stage.
Pi’s MCP Adapter Can Return Values Instead of Pages
Pi’s tool connector adds an optional Cloudflare reader for long replies. It sends the main agent named values, such as amounts and dates, plus a reference to the saved text.
OpenWork Adds Charts and 3D to Its Generated Apps
An agent-created interface can do more than display text and buttons. OpenWork has added a small catalogue of visual libraries it can bundle into an app—expanding what authors can build without asking them to install packages.
How AI Agents Work
The long-form map behind the daily paper: context, tools, loops, memory, delegation, safety, interface, and trust.