Daily Digest — 2026-09-28

Kevin Hou, Head of Product Engineering at Google Antigravity, presented on "Get Out of the Model's Way" at the AI Engineer World's Fair. He emphasized that Large Language Models (LLMs) are no longer just role-players but can be star players if the right product is built around them.

Google Antigravity is Google's agentic coding product for both technical and non-technical users, launched in November 2025. It's built with an "Agent-First" philosophy. The initial offering was the Antigravity IDE, featuring a new Agent Manager concept to manage and orchestrate multiple agents. With Antigravity 2.0, launched recently, the IDE was decoupled from the Agent Manager, making them two separate applications. This allows users to leverage the Agent Manager as a standalone app.

The philosophy of "Scaling with Intelligence" has guided Antigravity's development since 2022. This means that as the underlying models improve, the product's capabilities and user experience should also evolve. The frontier edge of the model's capabilities should be apparent in the user's product experience.

Hou illustrated this evolution through a timeline: 2022 (Autocomplete + Chat): Focused on deterministic features like embedding search, rules files, code snippets around cursor, AST parsing, and deterministic context retrieval. The models at this stage could only handle deterministic tasks. 2024 (Agents): As models improved, new primitives emerged, including agents with custom tools, MCP server integrations, and permissions. Agents enabled more dynamic workflows. 2025 (Agent Manager): This era saw the introduction of a platform to manage multiple agents in parallel. New primitives like skills & plugins, hooks, and artifacts (e.g., plan & walkthrough docs) facilitated agent-first management UIs. 2026 era (Today's focus): The current focus is on Agent Teams and Software Factories, where the Agent Manager and Traditional IDE are decoupled applications. The speaker believes the IDE is to the Agent Manager what the debugger was to the IDE – a specialized tool used when needed for deeper inspection rather than constant interaction.

Hou discussed two past challenges in "Scaling with Intelligence" and how they were overcome: 1. Giving AI a Terminal: Initially, there was fear (rightfully so) about the potential downsides of giving AI direct terminal access. However, as models improved and investment was made in permission primitives, users could build faster and more safely. Models became smarter at deciding what they should and shouldn't run. 2. Removing Chat, Only Agent: The decision to remove the familiar chat sidebar in favor of an agent-only interface faced significant user pushback. Despite initial resistance, multi-step, agentic research and execution proved to be the future, and users now embrace these agentic products.

For 2026, Antigravity is introducing Agent Teams in public preview. This feature allows users to invoke a team of agents to autonomously tackle large projects. The workflow involves a user specifying a task to a Lead Agent, which then manages a team of Subagents of arbitrary size to get the work done. Key aspects of Subagents include: Dynamic model selection: Subagents can select different models from the main agent. Parallel worktrees: Subagents can operate in parallel. Sandboxing: Secure execution environments. On-the-fly roles: Specialized roles for each subagent. * Dynamic prompting: Agents are prompted dynamically based on task needs.

Antigravity has already demonstrated impressive capabilities with Agent Teams: Example 1: OS Kernel with DOOM: Built an entire OS kernel from scratch and played DOOM on it. This involved 93 subagents, 12 hours of runtime, 15K+ model requests, 2.6 billion tokens, and cost less than $1K in API credits. This highlighted how subagent primitives are defining the 2026 era of agent teams, enabling complex, scalable, and affordable projects. Example 2: Side by Side Eval Analysis: Automated research tasks where multiple rollout results are compared. Traditionally, this involved extensive manual effort and Jupyter Notebook work. With Antigravity, researchers can automate 90% of this workflow by simply asking the agent questions in natural language. The agent then crunches numbers, identifies deltas, and proposes 100 different hypotheses for why those deltas occurred. It then spins up a subagent for each hypothesis to drill down in parallel, mapping back to a single response and providing a report. This process, previously taking hours, now takes minutes.

New primitives for 2026 being introduced include: Subagents: Dynamically configured subagents to assemble an agent team. They feature dynamic model selection, parallel worktrees, sandboxing, on-the-fly roles, and dynamic prompting. They enable asynchronous subagent execution and communication, leading to specialized, fully agent-driven roles. Sidecars: Long-lived utilities that the model can run and edit itself. This new plugin protocol enables SMS-based messaging, webhooks, cron jobs, other conversation triggers, review agents, and always-on personal assistants. Antigravity already uses sidecars for time-based tasks like scheduled cron jobs. * Generative UI: Very specialized UIs are considered dead. Agents are now lightning-fast, streaming in ~900 tokens/second on Antigravity (10x faster than other frontier models). This allows for dynamic generation of charts, dashboards, and one-off widgets. Instead of relying on fixed templates or HTML files, Antigravity can render generative UI inline, allowing users to go from a thought to a fully designed, interactive use case embedded in their conversation view within seconds.

The speaker concluded by quoting Steve Jobs: "They all have these keyboards that are there whether you need them or not to be there. And they all have these control buttons that are fixed in plastic and are the same for every application." Analogously, Antigravity aims to dynamically scale with the agent's needs, skipping heavy infrastructure and mechanical UIs in favor of sidecars and generative UI, creating a product experience that is not fixed in plastic. With the right primitives, the agent can do more than expected.

Meta's annual Connect event in 2026 centered around the introduction of Muse, a personal AI agent intended to operate across various user interfaces, including smart glasses, computers, email, and a new keychain-sized device called Muse Charm. Unlike previous Meta AI initiatives like Facebook M, Muse is powered by Muse Spark 1.3, a large language model that enables it to browse the web, interact with local applications, and manage emails. Meta is offering Muse to users for free, with the caveat that it will take a cut of purchases made through partner retailers, raising questions about monetization and user data.

A significant portion of the keynote addressed privacy concerns, with Meta emphasizing a novel architecture called Muse Secure VM. This system proposes an isolated Linux Virtual Machine for each user, containing the agent's browser, storage, and a Rust harness (Hatch). Passwords are kept separate from the agent's sandbox and accessed via a gatekeeper called Sentinel, which swaps in real tokens only for legitimate requests. This design aims to prevent even Meta from accessing sensitive user data directly. To demonstrate confidence, Meta announced a bug bounty program, offering up to $130,000 for successful prompt injection attempts. However, Meta's privacy policy clarifies that it does not prevent Meta from accessing data when necessary to support, secure, or operate the service, indicating a potential loophole. A future Muse Confidential VM is planned to prevent Meta from accessing data cryptographically, but this is currently in testing and not the default.

Hardware announcements included the new Meta VR Glasses, weighing 100 grams with 5K micro-OLED displays and IMAX Enhanced certification, priced at $1,300. These are slated for a Spring 2027 release. The existing Meta Smart Glasses (Ray-Ban Meta) also received an update (Gen 3), featuring a new design, an AI button, and a price point of $449. An audio-only version is available for $349. Notably, the presenter highlighted the appearance of Palmer Luckey, founder of Oculus (acquired by Facebook), who has previously been critical of Meta. His positive endorsement of the new VR glasses was presented as a significant validation, despite his previous public criticisms and his subsequent venture into defense technology with Anduril. News reports also surfaced just hours before the Connect presentation, alleging that Muse had accessed a tech columnist's private iMessages without permission, directly contradicting Meta's privacy claims and casting doubt on the immediate security assurances.

For developers, Meta opened up Muse Connectors, allowing third-party integration with the Muse platform, framed as a "generational opportunity" to build on a new platform. The video's sponsor, Hyperagent, showcased its own platform for building AI agents, emphasizing team collaboration through shared workspaces, agents, skills, and memories. Hyperagent allows agents to leverage various large language models (e.g., Opus 5.5, DeepSeek V4 Pro) and facilitates handoffs between AI agents and human team members, offering a more immediate and customizable solution compared to Meta's upcoming offerings.