Google's Antigravity adds a /boost deep-reasoning mode
Google's coding-agent IDE gets a one-command "think harder" pipeline that spins up subagents to split, verify, and rebuild hard problems, while CrowdStrike and Intel push a Falcon AI model down into the NPU on Dell's AI PCs to catch the sensitive data employees paste into chatbots.
Google is rolling out a new execution mode inside its Antigravity coding-agent IDE that breaks a hard engineering task into a three-phase multi-agent pipeline, dispatches the work to isolated subagents, and runs verification rounds before returning a final answer. The new mode — triggered by a single in-product command — splits problems into focused subtasks, hands each one to a subagent running in a clean, ephemeral worktree, and aggregates the results with regression checks before showing the user anything (Google Antigravity docs, AlphaSignal, Hacker News discussion). Each subagent inherits the workspace's file-access and command-permission rules, so protected edits still prompt the user for confirmation — the orchestration runs deeper, not looser. The feature slots between the default single-agent loop and Antigravity's heavier long-running team mode, and lands first on paid Pro and Ultra plans; the token cost is meaningfully higher because you're paying for orchestration plus multiple subagents plus verification rounds. The take: every serious coding-agent harness is racing to productize the pattern its power users were already hand-rolling — split the problem, run parallel exploration, verify against tests, iterate — and turning that into a single button is the first step toward making deep reasoning the default, not a power-user trick. The risk for Google is the same one every agent platform now shares: when the productized version of "think harder" gets good enough, the boring default loop starts to feel like a downgrade.
CrowdStrike and Intel are moving AI security below the operating system, putting a Falcon data-security model directly on Intel's NPU inside Dell AI PCs to inspect the sensitive material employees paste into chatbots and assistants. CrowdStrike's chief business officer Daniel Bernard framed the move as a response to AI becoming the new endpoint — "security has to operate where those interactions happen" — and the joint work combines Falcon's threat detection with Intel's on-device acceleration across CPUs, GPUs, and NPUs (CrowdStrike press release, Intel newsroom, SiliconANGLE). The timing is the story: a 2026 Cyberhaven report cited by SiliconANGLE puts nearly 40% of all data moving into AI tools as sensitive, and rules-based data-loss tools can't keep up with the volume. Pushing the classifier into the NPU keeps the data on the device, drops latency, and lets the endpoint catch a paste to a chatbot before the bytes leave the laptop. The harder question is what comes next — once a Falcon model lives on the NPU, it's a short hop to a Falcon model watching the local AI agent itself, which is where the real defensive moat will be built.
What to watch: how fast Antigravity's default loop converges with its /boost mode as the multi-agent harness gets cheaper, and whether the CrowdStrike-Intel NPU stack becomes the de facto security layer that every AI-PC OEM ships with.
Do you think on-device AI security will become a default requirement for enterprise PCs the way disk encryption did — or will it stay a niche add-on? Tell us in the comments.
