Weekly AI Roundup

What data from hundreds of organisations says about AI in software delivery, why refactoring cuts token costs, benchmarking tech stacks for agents, and a governed chain from conversation to code.

OpenAI’s DevDay announcements, a million agent skills, why context compounds, and agents moving into Slack.

Small decision models versus generative LLMs, how much context is the right amount, and a repository-native take on agent task tracking.

Spec-driven development for coding agents, agent sandboxing tools maturing, a marketplace for hiring humans raising new security questions, and Anthropic edges into project-management territory.