What Microsoft Should Steal Next From the Frontier Labs (And Why We Need It)

I work at Microsoft, and here’s the honest take: OpenAI, Anthropic, and Google have shipped three things in the past quarter that we should absolutely copy (or, let’s say, take inspiration from):

1. Eval-First Development (Anthropic, OpenAI):

OpenAI and Anthropic both now treat evals as first-class citizens—not just benchmarks, but inputs to the dev process. Every model change runs through a battery of real-world and adversarial evals, with failures automatically looping feedback into the pipeline. We need this across Azure AI and Copilot. Right now, too much is still manual. Make eval-driven gating the standard for all API and product releases—it’ll catch regressions before users do.

2. Agent-Oriented APIs (Anthropic, Google):

Claude Code and Google’s Gemini API let you hand off not just prompts, but tasks—the API manages planning, tool-use, even long context. We need to expand Azure AI’s orchestration primitives: let devs define goals, constraints, and safe toolsets, not just prompt strings. This would massively reduce glue code in Copilot extensions and Foundry apps.

3. Open Weights with Verifiable Lineage (Meta, Mistral):

Meta and Mistral are giving devs transparent open-weights models with clear data provenance. Microsoft should push for open, auditable models—especially in industries like health and finance, where black box AI just isn’t acceptable. We have the cloud trust stack; let’s extend it to models so users know exactly what they’re running.

In short: put evals, agentic APIs, and model transparency at the core of our stack. If we do, the next generation of Copilot and Azure AI will be something I’ll be damn proud to ship as an engineer.