For agents, the value is clearer still: structured JSON output, reusable commands and built-in skills that let models interact with Workspace data and actions without a custom integration layer.

MIT researchers developed Attention Matching, a KV cache compaction technique that compresses LLM memory by 50x in seconds — ...

The big headlines on this release are efficiency, with OpenAI reporting that GPT-5.4 uses far fewer tokens (47% fewer on some tasks) than its predecessors).

Google senior AI product manager Shubham Saboo has turned one of the thorniest problems in agent design into an open-source engineering exercise: persistent memory. This week, he published an ...

Databricks' KARL agent uses reinforcement learning to generalize across six enterprise search behaviors — the problem that breaks most RAG pipelines.

An open-source collaboration brings voice and vision AI directly onto consumer hardware, keeping sensitive data off the cloud LONDON--(BUSINESS WIRE) ...