AI Daily Digest · 2026-10-29
🔬 New AI Craft
1. Isolate test data as well as code for parallel agents
Microsoft's sample gives each coding task its own Cosmos DB emulator and versioned fixtures, then checks independent writes and cleanup. It rebuilds disposable test state without cloning production data, offering a reproducible pattern for parallel agent CI.
https://devblogs.microsoft.com/cosmosdb/agentic-ci-cd-with-azure-cosmos-db-isolated-environments-for-a-swarm-of-ai-agents
2. DAYJOB: Tax evaluates complete professional deliverables
Surge tests professional judgment across 50 tax-work environments and 1,789 files. Its reported leading pass rate is 10.8%, with success requiring at least 95% of rubric criteria. The benchmark highlights consequential omissions and complete deliverables alongside calculation accuracy.
https://surgehq.ai/blog/dayjob-tax
3. H-Net studies learned byte chunking under controlled training comparisons
Cartesia compares learned byte chunking with fixed tokenization while controlling data, architecture and training compute in its single-stage ladder. The reported gains concern pretraining efficiency, rather than measured serving speed. The improved architectures are still slated for later release.
https://www.cartesia.ai/blog/from-tokens-to-bytes-part-1
🛠️ Tools & Tips
1. GitHub Agentic Workflows turns Markdown into Actions automation
Markdown-defined reasoning tasks compile into standard Actions workflows, with read-only defaults, sandboxing and checks on proposed outputs. Agentic Workflows is now generally available; using its default Copilot CLI engine still requires a Copilot license.
https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available
⭐ Open Source Highlights
1. Local Code publishes its app source with integrated on-device tooling
The GPL-3.0 application combines inference with built-in coding runtimes. Local checkpoints support offline inference, while selecting a cloud provider sends prompts, file contents and tool output to that service. The app's license should not be treated as the license for downloaded model weights.
https://releases.drawthings.ai/p/local-code-by-draw-things-open-source
📰 Industry News
1. Anthropic expands AI for Science through partnerships with 20 foundations, institutions and research groups
🚀 Major Releases
1. Google announces Gemini 4 Carbon with limited access for cyber defenders and trusted partners
