Microsoft Build · VS Code

What we learned shipping VS Code weekly (without breaking everything)

Josh SpicerEngineer, VS Code

Pierce BogganPM, VS Code

The story is not “we adopted agents.” It is how the world’s largest dev tool changed its engineering system so we could ship weekly with confidence.

cadenceship weekly

constraintkeep confidence

changeevolve the system

section 01

You adopted agents. Now what?

consequence 01 · product surface

Successful agent adoption changed what VS Code could ship.

editor foundation <18 months AI-native editor

coreEdit, debug, terminal

platformExtension ecosystem

remoteCloud workspaces

tabInline suggestions

chat + editAsk for edits

agentAutonomous agents

model capability

Better models made agent work shippable.

Code survival by model

committed edited rejected

sonnet-4.576%

opus-4.686%

auto75%

haiku-4.578%

gpt-5.3-codex80%

opus-4.581%

gpt-4.155%

gpt-5-mini65%

sonnet-4.670%

gpt-5.2-codex74%

team impact

That translated into more product work getting done.

Total commits by month

2024 2025 2026

3.5K2.5K1.5K0

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Jan

Feb

Mar

Apr

release-note highlights

Six months of shipping would have sounded impossible pre-AI.

weekly release train

orchestration

Agents became a parallel work surface, not just a chat mode.

Agent window Remote agents Copilot CLI

validation

Agents got closer to the app, the browser, and the bug report.

Browser tabs as context Agents request browser access Background browsers Device emulation Screenshot to chat Rich issue reports with video

model control

Teams got knobs for cost, latency, quality, and where models run.

BYOK for Business/Enterprise Custom endpoint provider Provider-grouped model picker Token usage visibility Thinking effort controls

customization

Agent behavior became something teams could shape around their own workflows.

Customizations window Skills Hooks Plugins

safety + observability

Longer agent runs became easier to inspect, constrain, and debug.

Command risk assessment Sandbox network modes OpenTelemetry traces Grafana dashboards Structured agent signals

system pressure

More shipping created more issues and PRs to absorb.

Queues grew alongside output

Jan-Apr 2026

Open issues

+3.4K since January

JanFebMarApr

Open PRs

+850 since January

JanFebMarApr

cadence shift

After 10 years of monthly releases, VS Code moved to weekly releases in February.

AI was moving faster than monthly releases.

Model gains needed to reach users quickly.

Smaller batches were safer to ship.

the rest of the talk

AI changes how teams work, not just how they code.

01 Inner Loop

Agents change the daily build, validate, debug, and performance loop inside VS Code.

02 AI-Native Engineering Systems

Issue, PR, and observability systems have to keep up with AI-speed work.

03 Planning & Collaboration

When building gets faster, alignment, DRIs, and decision loops become the scarce resource.

section 02

The inner loop changed from writing code to validating behavior.

demo

Demo

vsc-bench · launch loop

Models are optimized before and after launch.

vsc-bench · before and after launch

Evaluation continues after the model ships.

vsc-bench · model tradeoffs

VSC-Bench are offline evaluations purpose built for VS Code.

section 03

AI-native teams need an AI-native engineering system.

demo

Demo

section 04

When building got faster, alignment became the bottleneck.

the old world

Monthly planning worked by reconciling leadership direction with IC-owned work.

top-down

Leads set the frame

Goals, product direction, release constraints, and the rough shape of the iteration.

bottoms-up

ICs brought the work

Bug fixes, debt, experiments, user pain, and the items closest to the code and community.

endgame

The plan got rationalized

During the last week of the month, the team reconciled the two inputs into a public iteration plan.

sample artifact

The plan was explicit, public, and organized by area.

microsoft/vscode#286040Iteration Plan for January 2026

Iteration Plan for January 2026

4-week iteration ship early February Endgame: Jan 26-30 top-level plan items

Workbench

• Integrated Browser V1
• Simple Browser improvements
• Update UI improvements

AI / Chat

• Create new sessions from chat input
• Improve model management UI
• Polish Agent Sessions UI

Agents / MCP

• Support model-specific tools
• Support MCP registry URL property
• Improve custom agents

NES / Completions

• Improve import NES speed
• Prefer NES over word suggestions
• Add diagnostics to completion prompt

failure mode

The planning model broke when the environment moved faster than the plan.

01

Pace of AI

Model capability, user expectations, and competitive pressure changed inside the iteration window.

02

Alignment load

More possible work meant more need to keep teams aligned on priorities, quality, and product intent.

03

Compressed lifecycle

Discovery, implementation, validation, and release started collapsing into the same operating window.

collapsed software lifecycle

When building gets cheaper, judgment becomes scarce.

The hard part shifts to choosing what matters, aligning people quickly, and shipping it well.

implementation cost down decision load up feedback loops compressed

01decide

02build

03validate

04ship

scarce skilljudgment

new operating cadence

We moved toward faster alignment loops instead of heavier plans.

• Temporary workstreams form around the highest-leverage outcomes.
• DRIs make ownership explicit while preserving broad collaboration.
• Weekly goals force tradeoffs while the facts are still fresh.

planning-loop.shrunning weekly

1

Align on outcome

Reset priorities against the latest product, model, and user reality.

2

Name the DRI

Make the decision-maker and owner visible before implementation fans out.

3

Ship small batches

Use weekly releases as a forcing function for learning and correction.

collaboration model

AI made building faster, so humans had to align more often.

before

• Monthly planning cadenceA monthly planning rhythm was enough when the work moved in longer arcs.
• Traditional code ownershipTeams formed around long-lived areas and the code they owned.
• Meetings stayed intentionally lightweightAsync updates and a few standups protected maker time.

after

• Weekly planning cadenceWeekly resets keep priorities aligned with the latest product, model, and user reality.
• Workstreams around outcomesTemporary workstreams swarm the highest-leverage AI opportunities, then reset.
• Meetings intentionally force alignmentDaily — sometimes twice-daily — syncs keep decisions, validation, and release risk moving together.

return process.evolve()

AI changes the work. Teams have to change how work works.

Thank you!