.NET team assigns bounded work to GitHub Copilot coding agent in dotnet/runtime: 878 agent PRs in ten months, 535 merged, with AI review on every PR (vendor's own product)
Vendor's own product: Microsoft's .NET team reports ten months (May 2025-March 2026) of using GitHub Copilot coding agent (CCA) in dotnet/runtime, the repository behind the .NET runtime, JIT, GC and core libraries. Maintainers explicitly assign bounded tasks; CCA works in a sandbox and opens a PR, and every PR, whether from CCA, staff or the community, also gets automatic Copilot Code Review. Early runs failed because the repo had no copilot-instructions file and the sandbox firewall blocked the NuGet feeds the build needs; the team added instructions and access first. Microsoft reports 878 CCA PRs (14% of all PRs), 535 merged, a 67.9% success rate versus 87.1% for Microsoft engineers' PRs, and 3 of 535 merged CCA PRs reverted (0.6%) versus 0.8% for other PRs (reported by Microsoft; the post says the data is directional, since tasks given to CCA were hand-picked and more bounded)
878 CCA PRs over ten months, 535 merged; 67.9% success rate vs 87.1% for Microsoft engineers' PRs Reported by Microsoft; vendor's own product, non-random task selection · Source (devblogs.microsoft.com)
3 of 535 merged CCA PRs reverted (0.6%) vs 33 of 4,251 other merged PRs (0.8%); the post says the difference is not statistically meaningful Reported by Microsoft; vendor's own product · Source (devblogs.microsoft.com)
1Password's engineers hand Codex well-defined user stories, which it breaks into functional specs and near-final prototypes for systems engineers to finish; Codex also reviews PRs before a human, runs acceptance tests in parallel worktrees, and investigates production issues by pulling evidence from incident management, telemetry, source control, paging and feature flags. Security is enforced by keeping secret references (not credentials) in repos and resolving them only when Codex calls an approved internal tool, plus AppSec policies packaged as reusable skills. OpenAI reports a 20.9% measured productivity improvement for the core cohort, a 10.9% cut in median PR cycle time, and investigation of a defect spanning 10+ microservices falling from about two hours to 5-20 minutes. 1Password's 553% ROI and ~$784K annual capacity value are modeled estimates (50 developers, $250K loaded cost, 40% 'directional' Codex attribution, 75% realization) (reported by OpenAI, citing 1Password)
AgentMail's AgentID gives agents email-based identities tied to a human owner so services like Context.dev can allow autonomous agent signups without fraud.
AMD uses GitHub Copilot coding agent to auto-triage 1M+ nightly simulation failures, propose fixes for morning review, and Autofix to resolve ~70% of vulnerability classes automatically, compressing feature cycles from weeks to days across 20,000+ engineers on GitHub Enterprise (reported by GitHub / AMD)
The ERP platform planning team at Korean beauty group AMOREPACIFIC (Sulwhasoo, Laneige and others) ran an August 2025-March 2026 project with AWS to bring AI-driven development to its SAP ECC 6.0 (NetWeaver 7.40) estate, where ABAP lives inside the SAP server and the SAPUI5 front end uses a custom NWDI portal framework. The team chose Kiro IDE: in Spec mode the agent turns a Jira ticket and FS/TS documents into requirements, design and task specs and then generates ABAP/SAPUI5 code; Vibe mode handles chat-driven changes to existing modules. An SAP MCP server built by AWS solutions architects gives the agent read access to RFC functions and table structures, and 'steering' files encode in-house coding conventions; developers review the generated code. Across four projects (one ABAP, three SAPUI5), the team reports productivity gains of up to 72% against an earlier 20-30% estimate for generic tools, and documentation time cut by 80% (reported on AWS's Korean tech blog; 'up to' figures from internal analysis over four projects)