9 articles · 3 categories
The finishable daily brief
Sunday, Oct 11, 2026
9 articles · 3 categories
read top to bottom · then stop
In 30 seconds
Context management moved into the model: Meta, MIT and the University of Washington introduced Context Language Models, which edit their own context instead of relying on predefined rules.
Builders also shipped coding-agent tooling and harness write-ups, while DeepSeek's V4.1 Flash reached a hosted inference API and DeepSeek and Huawei were reported shipping six Ascend tools aimed at CUDA.
Context handling and memory are being rebuilt as explicit, inspectable components, from model-managed context to plain-file memory and hand-built harnesses.
Meta, MIT and UW researchers propose models that edit their own context rather than follow predefined management rules, targeting better performance at lower compute.
Open-source project storing coding-agent long-term memory as a tree of plain files.
Walkthrough of building an agent harness from scratch.
Two new coding agents appeared: one that composes its own multi-agent workflows and one that runs in the browser.
Free coding agent that generates its own multi-agent workflows.
Coding agent that runs in the browser.
DeepSeek's V4.1 Flash gained a hosted API, and a report says DeepSeek and Huawei released six Ascend tools positioned against CUDA.
Dolphin's inference API went live serving DeepSeek V4.1 Flash.
Report of six DeepSeek and Huawei Ascend tools aimed at rivaling CUDA.
You are caught up for this edition