How VS Code Builds with AI
🎬 Watch The Story of VS Code! VS Code is now on Instagram, follow @vscode.ig for updates, tips, and more. Rewatch GitHub Copilot Day. How VS Code Builds with AI March 13, 2026 by Pierce Boggan We use AI every day to ship VS Code. It's made us so much faster that, after ten years of monthly releases, we just went weekly. Agents were the key that unlocked this, not just for writing code, but across every part of how the team works.
To kick off Agent Sessions Day , I sat down with Peng Lyu, Engineering Manager on the VS Code team, to walk through how the VS Code team actually uses AI for our day-to-day work. Not only for implementing features (that part's self-evident), but for everything around building features: triage, code review, release notes, validation, staying productive in a meeting-heavy schedule. In that session, we probably only covered 5% of what the team does with agents on any given day.
But it's representative of how a product that's used by millions of developers is built. So, we wanted to share more about these big recent changes to our workflow, and where we think we're going next. After Ten Years of Monthly Releases, We Went Weekly We shipped VS Code monthly for ten years. Every single month we went through our well-oiled cycle of plan, build, test, endgame, ship. With each member of the team rotating through the different roles, it was a rhythm that became part of the team's culture.
Recently, we decided to start shipping VS Code on a weekly cadence. And we wanted to keep the bar for rigor and quality just as high. A monthly cycle gives you breathing room with time to plan, time to run a full endgame week where the team cross-tests each other's features, and time to write thorough release notes. Moving to a weekly cadence means all of that has to get faster or get automated. This is a huge change and a year ago, we couldn't have done it.
This shift was only possible because of the way agents have transformed how we work. The weekly cadence isn't about shipping faster for its own sake. It's about getting improvements to developers sooner. A bug fix that used to wait three weeks for the next stable release now ships in days. A feature that's merged on Monday can be in developers' editors that same week. That feedback loop of ship > learn > iterate just gets so much faster.
What We Learned From This Shift Our workflows and processes continue to evolve daily, as we learn and adapt. But there are some key learnings that continue to hold true: Parallelize yourself. Build the habit of kicking off multiple agent sessions before context-switching. Worktrees, cloud agents, multiple VS Code sessions… use them all. Skip the intermediate artifacts. What used to be meeting notes → issues → specs → code, is now actually becoming meeting → agent sessions → code → PR.
Automate the overhead that scales with velocity. We built agent-powered pipelines for issue triage, commit summarization, release notes, code review, all of it using Copilot CLI, the Copilot SDK, and GitHub Actions. Engineers are still on the other end of these workflows, but agents help surface the right things to the right people, faster. Invest in harnesses before speed. Tests, golden scenarios, and code review gates prevent agent-driven velocity from becoming agent-driven regression.
Ownership is evolving. When PMs, engineers from other areas, community contributors, and agents can contribute to any component, traditional ownership models need to adapt. Accountability for outcomes still rests with engineers. Keep humans in the loop for taste. Agents check correctness. Humans evaluate delight. Let's look into more detail on how each of these plays out in our team. Working in Parallel There's a famous Paul Graham essay about how maker schedules and manager schedules are fundamentally incompatible.
That was mostly true until recently. But agents are changing that. Here's what a typical day might look like: Before entering a meeting, kick off 3-4 agent sessions for fixing bugs, prototyping features, or triaging issues. During the meeting, agents run in parallel across multiple VS Code sessions, worktrees, or in the cloud. After the meeting, review the agent output, verify locally, merge or re-prompt, and kick off again.
Managers still attend meetings and other managerial tasks, but they can use agents to also take on some of the maker work that used to be impossible to do in a meeting-heavy schedule. Let me give you a real example. Peng starts each morning by updating VS Code Insiders . Most days, we ship Insiders builds twice a day so we can get early feedback on the stuff we are working on. Then he runs a custom agent that fetches his meetings via Work IQ , and produces a snapshot of what's on his plate.
From there, Peng decides what needs his focus, what to delegate to agents, and what to prioritize for the team. The agent handles the busywork of gathering context so he can cut straight to the interesting problems. By the time he's in his first call, tasks are already running in parallel. "Previously, you were always working sequentially. You wrote notes, turned them into issues, and then someone else, or you, would pick that up later.
Now you are empowered and able to do things in parallel. It's a habit you have to build. So, I don't write down meeting notes anymore. I'm kicking off the agents directly." — Peng Lyu It really is a new muscle. Someone in a meeting mentions something we need to go do, and I'll fire off agents right there. We’ve also enabled transcription for most of our meetings in Teams, so grabbing context after the fact is easy.
What used to be meeting notes turned into issues turned into work is now just a prompt kicked off in the moment. Automating the Overhead More velocity is great. It also creates its own overhead: more issues to triage, more commits to track, more release notes to write. Here's how we've automated the parts that scale with speed. Commit summarization. We built a custom slash command that fetches all commits from the last 24 hours across multiple repos and summarizes them with a fast model.
It used to be you'd git fetch and have 20 or 30 commits. Now, there can be 100+ waiting for you. An entire feature area can land in a single day. That same pipeline feeds into our Insiders changelog and powers our automated X account that posts daily updates. All built on Copilot CLI and the Copilot SDK, running as GitHub Actions triggered by commits to main. Issue triage. VS Code is one of the largest open-source projects on GitHub.
We love our community, and the volume of issues we get is a reflection of how many people care about the product: hundreds land daily. We used to have a rotating "inbox tracker" role, one person triaging everything for a week. This no longer scales. Now, every time an issue is opened, it triggers an agent loop in GitHub Actions that detects duplicates (with confidence scores), determines the right owner, and suggests labels.
The agent reads our ownership docs and looks at historical assignment patterns, because ownership shifts over time. You can see it in the repo's public data: comparing Jan-Mar year over year, commit volume has more than doubled and the team is closing nearly 3x as many issues. Better triage helps engineers find and fix the right issues faster, which frees up more time for actual software development. "Now that piece of code is written by Copilot, who is the right owner for it?
I would say it's still our engineers who are accountable for the outcome. But you do need the right harness to welcome other people to contribute to your component." — Peng Lyu The team also built a Chrome extension that shows triage suggestions directly on GitHub issues, like duplicates, owners, and labels. It includes a dashboard showing issue status across the team. Inside VS Code, custom slash commands let engineers groom issues and find duplicates without leaving the editor.
We've already seen a real boo
The rest of this page is on the live site.