Microsoft's VS Code team has done something significant: after a decade of monthly releases, they have switched to a weekly release cadence.
Not because they hired more people, but because AI agents fundamentally increased the team's delivery speed. In a March blog post, they shared concrete numbers — perhaps the most direct evidence of AI coding effectiveness I have seen.
Hard numbers
Year-over-year (January–March):
- Code commits: 2,339 → 5,104, a 2.2x increase
- Issues closed: 2,916 → 8,402, a 2.9x increase
The same team, the same people, using AI agents — output doubled. No mass hiring, no outsourcing. Just tools driving efficiency.
What they actually changed
The VS Code team summarized six lessons, several of which are particularly noteworthy:
Skipping intermediate steps
The old process: meeting → meeting notes → break down issues → write requirements → write code → submit PR.
The new process: meeting → launch an agent session → code directly → PR.
All the middle steps are eliminated because the agent can directly turn meeting outcomes into implementations, without requiring a human to translate requirements.
Running multiple agent sessions in parallel
Instead of having AI do one thing at a time, they launch multiple agent sessions during context-switching gaps, advancing several workflows simultaneously. This is enabled by worktrees and cloud-based agents.
Automated pipelines handle scale burdens
Issue triage, commit summary generation, release notes, code review — all these repetitive but necessary tasks are now handled by dedicated pipelines built with Copilot CLI and GitHub Actions. Machines do the work; humans don't touch it.
Product managers skip requirements docs, write PRs directly
This change is somewhat disruptive. PMs now validate ideas by directly implementing prototypes rather than writing specification documents. Building a feature and then iterating is much faster than describing it and waiting for someone else to build it.
Build quality assurance first
With increased speed comes higher regression risk. The VS Code team's approach is to invest upfront in test infrastructure, golden scenario documentation, and code review gates — balancing speed and quality through upfront engineering, not hope.
Humans judge "goodness," agents judge "correctness"
This quote hits the mark: "Agents check correctness; humans judge the experience."
Automation can ensure code runs, but whether the user experience is good and whether the interaction feels natural — that judgment still requires a human.
What's new in v1.115
The April release of VS Code 1.115 continues to strengthen the agent-native experience:
- VS Code Agents App (Preview): A companion application designed for agent-native development, enabling parallel processing of multiple project tasks and cross-project monitoring and review of sessions
- GitHub Copilot BYOK support: Enterprise users can now bring their own model API keys instead of being forced to use Microsoft's default models
- Tools: Agents can now send input directly to background processes, handling interactive CLI programs
A notable point
There is something special about the VS Code case: they are writing about developing VS Code itself. AI agents are helping build the platform on which AI tools run — a self-accelerating process.
More interestingly, their experience contains no story of "AI replacing a role." Instead, PMs are writing PRs, engineers are advancing more tasks in parallel, and everyone's scope of work is expanding. The productivity improvement manifests as doing more, not as having fewer people.
This may be the mainstream impact of AI coding tools in the short to medium term: not replacement, but multiplication.
At least, that is the case in the VS Code team's story.
Sources: How VS Code Builds with AI (code.visualstudio.com); CocoLoop, Visual Studio Code April 2026 Release Notes (Releasebot)