Microsoft AI Updates: November 3, 2026

1. Microsoft Details How Nine Engineers Built Copilot Cowork With Agents, Then Moved Its Live Runtime From Python to Rust

Microsoft partner software engineer Bas Brekelmans described how a small team took Copilot Cowork, an agentic execution layer that runs multi-step workflows for Microsoft 365 users, from a first commit on January 13 to Frontier customers on March 30 and worldwide general availability on June 16, with agents writing most of the code; on October 21, nine engineers merged 1,211 pull requests, and the week of October 12 saw 4,581 merges. The post lists six practices: a markdown spec before any code (what the change does, how to know it works, what it must not break); narrow agents in separate repo copies, with the model running in a container without user credentials while a trusted host owns identity, policy, approvals, tools, and audit; competing reviewer bots (median first review about four and a half minutes) backed by a human check on critical changes; checks that report what they checked and ship with a test that makes them fail on purpose, after a size-limit check piped through tail stayed green even though the limit had been exceeded; about 30 rule files in the repo; and strict crate roles (contract, kit, implementation, binary) that took implementation-to-implementation dependency exceptions from 216 to zero and the longest sequential rebuild chain from 11 crates to 3. Automated loops run with a written keep-or-revert rule and attempt limits: a test-speed harness merged about 300 PRs that cut the touched tests from roughly 17,700 to 1,700 seconds of CI time, and a “factory” harness turns production signals into work items, uses critics from a different model family, and reverts each fix to confirm the new tests fail without it. For the Rust move, the team recorded about 430 fixtures from the live Python code for Rust tests to replay, shifted traffic one cluster at a time from September 23 to October 10, and deleted the Python runtime (about 8,000 files and 4.9 million lines, including tests and generated code); first-turn time to first token fell from 7.9 to 2.7 seconds at p50 and from 14.5 to 6.2 seconds at p90, which the author attributes partly to moving the agent loop out of the container, and the team makes no cost claim yet. Source

2. Microsoft and Lumen Expand Co-Selling of AI-Ready Enterprise Networks for Azure

Microsoft and Lumen Technologies announced an expanded co-selling collaboration aimed at connecting the data, applications, and infrastructure behind enterprise AI workloads across public clouds, private data centers, and offices. Lumen intends to bring its full suite of cloud networking and application connectivity capabilities to Azure, and the companies plan joint account planning where network modernization, Azure migration, and AI projects overlap; Alkira, a Lumen Connect solution, is already sold through Microsoft Marketplace, with more Lumen capabilities to be added there over time. The companies also plan to explore how Lumen’s connectivity works with Microsoft’s agent governance and security capabilities, building on a relationship in which Lumen’s Private Connectivity Fabric already supports Microsoft data center connectivity. No pricing, product integrations, or timelines were given, and both companies say details will follow as joint go-to-market plans progress. Source