What Microsoft Needs to Borrow from OpenAI and Anthropic Right Now

Let’s talk brass tacks: the Copilot family (VS Code, GitHub, M365, even Azure OpenAI Studio) is good—but it’s missing key pieces that OpenAI and Anthropic are already betting on at scale. Here’s what Microsoft needs to steal, yesterday:

1. First-Class Eval Loops

OpenAI’s latest developer tools make eval-driven iteration seamless. You write evals as code, bundle them with your agent, and get continuous pass/fail metrics as you ship. Anthropic’s Claude Code brings evals so close to the dev loop, you can’t not use them. Microsoft’s current Copilot eval story? Way too bolted-on. We need eval suites built-in at every abstraction—function, agent, UI—and surfaced as first-class CI/CD signals. This would help teams catch regressions before prod, not after the user tweets a bug report.

2. Real Agentic DevTooling

Anthropic’s deep integration of loop engineering and structured specs into Claude Code is a model Microsoft should copy, fast. GitHub Copilot should let you compose plan→act→verify→fix loops, not just single-shot completions. Imagine designing loop templates, attaching unit/spec/test checks, and running them as part of the Copilot workflow. This is how you get from autocomplete to actual agentic coding.

3. Native Tool Server Integration

OpenAI’s tool calling and Anthropic’s MCP servers have set a new standard: let LLMs call real code, real services, real-time as first-class citizens. Microsoft should embed MCP/server-style action routing directly into Copilot, Azure ML, and even Windows DevHome, so agents don’t just suggest—they act. This closes the loop from suggestion to execution, and it’s how the next wave of dev tooling will be built.

The ask isn’t new features, it’s about closing the workflow gap. If Microsoft moves fast on evals, agentic loops, and tool servers, Copilot and Azure can keep leading. If not, it’s going to get leapfrogged, fast.