Daily Agentic Field Watch - 2026-10-07 21:00 UTC
- Microsoft’s Oct. 7 Copilot Studio update makes apps, deterministic workflows, and agents composable in one platform: apps are in public preview, GitHub Copilot-harness hooks can enforce or block tool calls at lifecycle points, and Foundry IQ grounding is generally available. This is a notable enterprise shift from chat-agent building toward full process applications with control and governance; claims are Microsoft-described. Source: https://www.microsoft.com/en-us/copilot/blog/copilot-studio/new-and-improved-build-apps-extend-agents-and-transform-business-processes-in-microsoft-copilot-studio/
- Liquid AI released open d1 decision models: d1-3B handles text/images, while experimental d1-omni-600M takes text with images or audio. Unlike chat models, they return probabilities for typed decisions; local weights make low-latency routing, classification, and multimodal agent perception possible without a generative call. Liquid reports d1-3B averaging 74.1 on 11 image benchmarks; this is vendor/model-card evidence, not yet independent validation. The decision-model category was already gaining HN discussion, but today’s release makes the edge/open-weight angle concrete. Sources: https://www.liquid.ai/news ; https://huggingface.co/LiquidAI/d1-3b ; https://huggingface.co/LiquidAI/d1-omni-600m ; https://www.reddit.com/r/LocalLLaMA/comments/1wzuto0/new_lfm_to_be_released_today/ ; https://news.ycombinator.com/item?id=49984025