VS Code 1.120 dropped today, and if you're still treating AI agents like a novelty feature tucked in a sidebar, you're behind. Microsoft isn't hedging anymore—agents are the primary development interface now, and this release makes that crystal clear.
The release notes read less like "here's a cool AI thing" and more like "here's how we're rebuilding the editor around autonomous agents." Claude agent gets full planning support. BYOK users can tune reasoning effort. Password prompts in terminals are intercepted before they leak into chat logs. This isn't polish—it's infrastructure.
Planning Goes Multi-Agent
The headline feature: Claude agent now supports planning, matching what Copilot CLI has had for weeks. You give Claude a complex task, it breaks it into steps, shows you the plan inline, and you can edit or comment on individual steps before it executes.
This matters because it closes the capability gap between Microsoft's first-party Copilot CLI and third-party agent integrations. If you've standardized on Claude (or any other model via BYOK), you're no longer stuck with a second-class agent experience. The UX is identical—same plan editor, same inline feedback loop.
The editing part is equally critical. You can now provide feedback on individual plan steps inline without breaking the conversation flow. Agent proposes a step you don't like? Click it, revise it, keep moving. No more "stop, regenerate the whole plan, hope it's better this time."
Context Size Picker
The context size picker lands in the model picker dropdown. You can now explicitly choose 8K, 32K, 128K, or 1M context windows depending on your task and token budget.
Why this exists: most developers don't want to think about context sizes until they hit a limit or burn through credits too fast. But once you're running agents in production workflows—reviewing PRs, refactoring modules, analyzing logs—the ability to dial context size up or down based on the job becomes a cost and performance lever.
This also pairs well with the chat.tools.compressOutput.enabled setting that shipped in the same release. Compress tool output to reduce token usage without sacrificing agent functionality. Combined with smart context sizing, you can run longer sessions without hitting token caps or bleeding budget.
Terminal Secret Detection
Agents now detect password prompts in terminal output—sudo, ssh, gh auth login, and others. When a prompt is detected, VS Code surfaces a "Focus terminal" button and announces the prompt to screen readers. Secrets never route through the model.
This is one of those unglamorous features that prevents catastrophic mistakes. Without it, an agent running a deploy script could encounter a password prompt, get stuck, and either fail silently or worse—send the prompt (and your attempted workaround) straight into the chat context where it gets logged.
It's also a forcing function for better agent design. If your automation relies on interactive password prompts, that's a smell. This feature makes that smell visible and gives you a manual escape hatch while you fix the underlying issue (environment variables, SSH keys, credential managers).
BYOK Gets Reasoning Effort Tuning
If you're running bring-your-own-key models, you can now specify reasoning effort per request. This is the same parameter that Claude and GPT-5 expose to control how much "thinking time" the model spends before responding.
Low effort: faster responses, good enough for simple refactors or documentation questions. High effort: slower, deeper reasoning for complex debugging or architectural decisions. Previously you were stuck with the model's default, which wastes time on easy questions and shortcuts hard ones.
This feature is mostly relevant if you're paying per-token for reasoning models. The ability to tune effort per task can cut costs significantly when you're running dozens of agent sessions per day. It also signals that Microsoft expects BYOK to be a mainstream deployment path, not a niche escape hatch for hobbyists.
The Weekly Cadence Continues
VS Code moved to weekly stable releases in April, and 1.120 is the third weekly drop. The velocity is real: planning for Claude, context size picker, terminal secret detection, and compressed tool output all shipped in a single week.
That pace only works if the underlying architecture is solid. The fact that Microsoft can ship agent features this fast without breaking existing workflows suggests they've cracked the abstraction layers—agents are a first-class primitive now, not a bolt-on extension.
It also means if you're still evaluating whether to adopt agent workflows, the window is closing. The tooling is moving faster than most teams can absorb. As I wrote about context engineering, the teams that win are the ones who treat AI tooling as infrastructure, not a novelty.
What This Signals
Three things are now obvious from the release cadence and feature mix:
Agents are the interface. Microsoft isn't building "AI features for VS Code." They're rebuilding VS Code around agents. The distinction matters.
BYOK is production-ready. Reasoning effort tuning, cost controls, and multi-model parity mean Microsoft expects enterprises to run their own models, not just consume Copilot.
Weekly releases are sustainable. If they can ship this much in a week without breaking things, the architecture is mature. Expect this pace to continue.
If you're not running agent workflows yet, 1.120 is a good entry point. Planning support is now model-agnostic, token management is better, and the sharp edges around terminal interactions are smoothed out. The tooling is ready. The question is whether your team is.
Top comments (0)