← Back to blog
AiAbout 6 min read

Google Moved Coding Agents Off Your Laptop. The Permission Prompt Is Still the Bottleneck.

Published Nov 7, 2026
Google Moved Coding Agents Off Your Laptop. The Permission Prompt Is Still the Bottleneck.

The Antigravity team's own description of the feature is the best summary: your tasks run in independent cloud virtual machines, so you fire off a task, step away for coffee, unblock permissions from your phone, and come back to passing tests and a clean pull request.

Version 2.29 started rolling out on November 3 and the team confirmed remote agents publicly on November 6. The changelog lists 22 changes, and the two that matter are remote agents and a voice mode built on Gemini Live.

The shift is easy to underrate. Coding agents have been local software since the category existed. They read your files, run your tests, use your tools, and stop when you close the lid. Moving execution into a managed VM changes what the product is, and it also exposes which parts of an agent's capability were never in the model at all.

Remote agents, and what they are not

Antigravity already had a feature called Remote Control, which drives agents running on your own workstation from another device. That still requires the workstation to be online. Remote agents are a different thing: sessions and tools execute in an isolated Google-managed environment, so the work continues when your laptop sleeps, your browser closes, or you switch from desktop to phone.

You can start a session from the desktop app, from the web at antigravity.google.com, or from the CLI with `agy --remote-agent`, which needs version 1.4.12 or later.

A remote project attaches resources. GitHub repositories are cloned and mounted through OAuth. Google Drive files are read on demand. You can upload files directly, up to 8 MiB each, 50 files, 100 MiB per project and 2 GiB per account. Google says credentials are encrypted at rest. Plugins for GitHub, Drive, Docs, Sheets, Slides and Calendar connect over OAuth.

Those upload limits are small in a way that shapes what the feature is for. Eight mebibytes per file is fine for a script and useless for a dataset. This is an agent for repository work and document work, not for data engineering.

The eligibility line is the interesting line

Remote agents are available on Google AI Pro and Ultra personal accounts. Not the free tier. Not Google AI Plus. Not Workspace or Google Cloud business accounts.

That last exclusion is odd for a feature whose obvious customer is a development team. A team plan would want shared projects, an audit trail, and an admin who can revoke access. Google shipped the consumer version first, which tells you the company is optimising for individual developers trying the product rather than for procurement.

It also means the "come back to a clean pull request" workflow is, for now, a personal workflow. Your agent works in a Google-managed VM under your personal subscription, against a repository you authorized with your own GitHub account.

What you lose in the cloud

The list of features unavailable in remote mode is the most informative part of the documentation: print mode, code search, `/diff`, `/rewind`, voice dictation, and locally installed MCP servers, plugins and skills.

Read that last group again. MCP servers, plugins and skills are a large share of what makes a coding agent useful in a specific company. The internal API schema, the deploy script, the runbook, the style guide, the migration checklist. All of it lives locally, installed by the developer, often pointed at internal endpoints the cloud VM cannot reach.

So the feature ships with a clean split. A cloud agent gets the repository, the public internet, and the model's general competence. It does not get your local tooling. For greenfield work that is fine. For the messy repository that needs the internal build wrapper to compile, the agent will hit a wall that has nothing to do with intelligence.

This is a transitional problem and it will get solved, probably by letting teams publish approved MCP servers and skills into a remote project. Until then, the honest description of remote agents is: excellent for tasks that only need code and a shell, limited for tasks that need your machine's setup.

Voice as a permission channel

Voice mode runs on Gemini Live, which hears your voice directly rather than a transcript, so you can interrupt mid-sentence and it adapts. You can plan, steer changes, and approve permissions out loud. Voice and playback speed are configurable, and the mode is limited to local conversations.

The detail that matters is the carve-out: routine permissions can be approved by voice, destructive actions cannot.

That is a sensible line and it reveals where the friction actually is. An agent running for forty minutes will stop several times to ask whether it may run a command, install a package, or delete a file. Each stop is a context switch for the human. Approving by voice removes the context switch for the boring ones while keeping a physical action for the ones that can hurt.

The stated goal is that you talk to the agent the way you would talk to a teammate on a call. Speech recognition is the easy part. A teammate on a call can be told "use Postgres instead" and will re-plan, while an agent needs the instruction to arrive at the right moment in its loop.

The rest of the week moved the same direction

Cursor shipped Build Live Apps, which takes a description and produces a hosted web app with a Postgres database, authentication and security scans, deployed behind a shareable link with SSO. Cursor says its own teams use it for dashboards, internal ops tools and onboarding portals. The interesting consequence is that the deploy step stops being a step.

LangChain added async subagents to Deep Agents, with five tools for starting, checking, updating, cancelling and listing background tasks. An update to a running task can either interrupt and restart it, which suits a correction like "use Postgres instead", or enqueue until the current run finishes, which suits an addition like "run the tests when you're done". Unfinished tasks are listed in the orchestrator's system prompt, and a finished subagent either gets picked up before the orchestrator's next model call or wakes it with a new run.

Three separate teams, three implementations of the same idea: the agent becomes a service that runs continuously, and the human becomes an approver who drops in and out.

The plumbing is close to solved

Put the pieces together and the architecture is visible. Execution in cloud VMs that outlive your hardware. Subagent orchestration with explicit interrupt and enqueue semantics. Deployment as a side effect of generation. Voice as an approval channel. A phone as the control surface.

What is not solved is the thing all of this is designed around: the permission prompt. Every one of these systems still assumes a human decides, dozens of times a day, whether an autonomous process may take an irreversible action. Google's answer is to make approval cheap and keep it necessary for destructive operations. That is the right default and it is also an admission that nobody trusts the agent to make that call.

The next version of this problem is a policy layer: rules that say what an agent may do without asking, scoped to a repository, an environment, and a time window, with an audit log. Until that exists, "fire off a task and step away" works exactly as long as nothing unusual happens, which is another way of saying it works in demos.

Related articles