# The Agent Just Stopped Living in the Chat Window

> Meta put an agentic model on a single consumer GPU, while xAI and OpenAI gave their agents computers in the cloud. Neither yet gives enterprises an agent identity they can govern.

- URL: https://dperez.com/posts/the-agent-just-stopped-living-in-the-chat-window/
- Author: David Perez
- Published: 2026-10-03
- Tags: agents, models, identity

On 10 August 2026, Meta shipped a 30-billion-parameter model built to run on a single consumer GPU, offline. The next day, xAI shipped the opposite bet: an agent with [its own computer in the cloud](https://x.ai/news/introducing-grok-bot), one that never has to touch your device at all. Seven weeks later, OpenAI shipped the same bet again, under a different name.

A chat-era agent lives behind a chat window, and closing the tab kills it. Meta's new model is a local agent: it lives on your own hardware instead. xAI's and OpenAI's new agents are cloud agents: each gets a computer of its own, so it keeps working after you close the laptop. None of the three is yet what an enterprise actually needs: an agent with an identity a company can govern, suspend, and retire the way it would an employee's.

## Shrinking the model until it fits in your pocket

Muse Glimmer is [Meta's answer: shrink the model until it fits on one device](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model). It has 30 billion parameters total, split into a 28-billion-parameter text decoder and a 2-billion-parameter vision encoder Meta calls a Perception Encoder. Meta built it by distilling a larger model, Muse Spark, down into Glimmer, then carried it through further training on long, agent-heavy tasks and a final round of reinforcement learning. [Its 52 layers mix cheap local attention with occasional full-context layers](https://huggingface.co/blog/muse-glimmer), and each key-value head is shared across sixteen query heads, a ratio Meta says cuts the KV-cache memory footprint by the same amount.

All of that engineering has one purpose. It fits the model onto hardware a person actually owns. At full precision, Meta says, 30 billion parameters needs upwards of 55 gigabytes of memory. No consumer GPU carries that much. Quantized to roughly 4-bit precision, the same weights shrink to under 20 gigabytes, which leaves enough headroom inside a 24- or 32-gigabyte card for working memory, the vision encoder, and a speculative decoding drafter called DFlash, all running at once. Meta benchmarked the result against Gemma4-31B and Qwen3.6-27B, though that comparison is Meta's own vendor-reported evaluation, not an independently reproduced one.

## Giving the agent a computer instead

Meta's fix changes where the model runs and how big it is. xAI's and OpenAI's fix leaves the model alone and changes something else: the environment the agent executes in. OpenAI calls its new agent a dot. A dot runs on its own cloud computer and its own browser, powered by a model OpenAI calls [GPT-6 Astra](https://openai.com/index/introducing-dots/), and reaches into a plugin ecosystem of more than 4,000 apps. xAI calls its version a Grok Bot, and it does the same job: a persistent machine in the cloud that can sign into a tool with no clean API and finish the work there, the way a person would. Either way, the agent's presence now outlasts the browser tab or chat window a person would otherwise have to keep open.

In xAI's own launch post, a product team member named Roman put the actual sell bluntly: _"There is a huge difference between 90% done and 100% done...Grok Bot can finish the swing, because the work lands where a human would put it, in the actual tool."_ That is the pitch, and reasoning quality has little to do with it. The differentiator is whether the thing can click the same buttons a person clicks, inside software that was never built with an API for a model to call.

Both companies are already describing fleets, not single agents. xAI's staff run a chief-of-staff Bot that coordinates specialists for the inbox, expenses, recruiting, and bug fixes, passing work between them in shared threads. OpenAI's own launch post says it expects teams of dots working together on a person's behalf, and is previewing specialist dots with their own identity, credentials, and access, aimed at enterprise work like invoice processing and contract review.

## The infrastructure question that actually matters

An agent with standing access to a CRM or a shared mailbox needs the same six answers a new employee needs. Who issued its credentials? What can it actually reach? Which of its actions need a human sign-off before they fire? Is there a log of what it did and when? Who can suspend it the instant something looks wrong? And when it's let go, does its access die with it?

OpenAI has already built partial answers into dots itself. Its own launch post describes [access controls an admin sets per connected app](https://openai.com/index/introducing-dots/), Custom Rules that decide which actions need approval versus which the agent can take on its own, and an Activity View that logs background work for a person to review. Specialist dots go a step further: OpenAI says each one gets its own identity and credentials, built for a company to assign to one bounded job rather than hand to a person. The question an enterprise buyer is actually weighing this quarter is whether a vendor's agent has permissions and an audit trail solid enough to earn that kind of standing access at all, not which model scores two points higher on a coding benchmark.

None of that answers who is accountable for what the agent actually did with that access while everyone was asleep.