Three things shipped or surfaced in the last two weeks that are worth your attention, and none of them is another benchmark chart.

xAI launched Grok Bot, always-on AI teammates that get a cloud computer, sign into your existing tools with your credentials, and keep working after you close your laptop. Meta's unreleased Muse Video model turned up in closed beta with native audio, and the early outputs are genuinely strong. And Block released Buzz, a Slack replacement where every AI agent carries its own cryptographic identity rather than borrowing a human's login.

One is a product you can pay for today. One is a model nobody outside a partner list can touch. One is free, open source, and not ready. Here is what is real in each.

Grok Bot entered public beta on August 11. The pitch is agents you hand actual jobs to rather than prompts you babysit. Each Bot works through multi-step tasks end to end while you are away, surfacing only when something needs approval, and produces output inside the tool being used rather than in a chat window.

xAI's internal examples are the most useful part of the announcement. A sales Bot researches accounts overnight, scores contacts for intent, drafts emails in each seller's voice, and leaves a queue for approval. A demo-readiness Bot checks the environment overnight, fixes broken seeds and stale data, and drops a checklist before morning calls. Others keep CRM records clean, process invoices arriving in Gmail, seat new hires, and reproduce bugs in the UI before filing tickets and handing fixes to a debugging Bot.

Access is gated behind SuperGrok Heavy at $300 a month, Cursor Ultra at $200, or Cursor Teams Premium at $120 per seat. Desktop including a Linux build, plus iOS. Android is listed as coming soon with no date. There is no free tier.

Now the part worth flagging. The launch page says Bots have their own computer. xAI's documentation, published the same day, says every Bot on your account shares one persistent cloud computer, and warns twice not to treat separate Bots as a security boundary.

That gap matters. It is the difference between giving your bookkeeping Bot and your inbox Bot separate credentials with a contained blast radius, and everything every Bot can reach living in one shared place. For a product whose entire pitch is signing into your tools and working unsupervised, the security model is the product. Credit to xAI's docs team for writing the sharp edges down. Someone should fix the marketing page to match.

Meta's video model is real and being tested. TestingCatalog obtained access to Muse Video, currently in closed beta with selected partners, and published the first outputs anyone outside that list has seen.

The assessment from people who actually ran it: state-of-the-art potential, particularly in fine detail, world understanding, and temporal consistency. It currently produces 10-second clips, with output quality described as remarkably high for that length. Native audio is confirmed, with voice and music support.

Meta first previewed Muse Video in July alongside Muse Image, promising availability to creators and Meta AI. What is new is evidence that development has moved past a public preview into working beta.

Two honest limits. Meta has previously acknowledged remaining gaps around audio-video synchronization and physically accurate fast motion, which are the two hardest problems in generated video and the ones that separate impressive demos from usable footage. And the model is already quite restrictive on copyrighted content, which will either read as responsible or frustrating depending on what you were hoping to make.

The strategic logic is hard to argue with. Facebook and Instagram already run on visual content, so Meta has distribution for a video model that nobody else can match. If Muse Video is competitive at scale, it does not need to win a benchmark to matter.

Block released Buzz on July 21, an open-source workspace built on the decentralized Nostr protocol, pitched as a replacement for both Slack and GitHub. It picked up 15,600 GitHub stars in its first week.

On the surface it is familiar: channels, threads, DMs, voice, media, code repositories, workflows. The difference is underneath. Every participant, human or agent, holds their own cryptographic keypair. Open an agent's profile and its public key sits next to yours. The member count does not say three people and one bot. It says four.

The practical consequence is the part that matters. Agents like Claude Code, Codex, and Block's own goose framework do not operate behind a human's login with a shared bot token. Each has its own identity, permissions, and audit trail, which means you can revoke one agent without breaking every other automation you run. Every message, reaction, code patch, and approval is a cryptographically signed event in one shared log, and because the identity belongs to the keyholder rather than the platform, it is portable across any Nostr-compatible system.

Honest caveats, and Block would say the same: no mobile clients, no SOC 2, and a single-relay architecture despite the decentralized framing. Messages on Block-hosted relays are not end-to-end encrypted, and agent providers may receive prompts and channel content. It is Apache 2.0 and self-hostable if that matters to you. It is not replacing your Slack this quarter.

Line them up and you get three theories about where the value is going.

xAI is betting on labor. Grok Bot is not sold as a better model, it is sold as a teammate you pay a salary-shaped price for. At $300 a month the comparison is not to an API bill, it is to a contractor.

Meta is betting on distribution. Muse Video does not need to beat every rival on quality if it ships inside Instagram and Facebook, where the audience for generated video already is. Meta has lost the coding-agent race and the frontier-model race this year. Video is the one where its existing surfaces are an actual advantage.

Block is betting on infrastructure. Buzz makes no claim to have the best model or the best agent. It is arguing that whoever wins those races, the identity and permission layer underneath will need to be open, portable, and owned by you rather than a vendor.

The uncomfortable common thread is that all three assume agents will be doing real work with real access, and only one of them has a serious answer for what happens when that goes wrong.

The stories moving fast beyond today's headlines:

Grok Bot. xAI's public beta launched August 11: always-on AI teammates that get a cloud computer with browser, filesystem, and terminal, sign into your tools with your credentials, and finish multi-step jobs unsupervised, surfacing only for approvals. Gated behind SuperGrok Heavy at $300 a month, Cursor Ultra at $200, or Cursor Teams Premium at $120 per seat, on desktop and iOS with Android unscheduled. The catch: the launch page says Bots have their own computer, while xAI's own docs say all your Bots share one and warn twice not to treat separate Bots as a security boundary.

Muse Video. Meta's unreleased video model is in closed beta with selected partners, and TestingCatalog published the first outputs from outside that list. It produces 10-second clips with native audio including voice and music, and testers describe state-of-the-art potential in fine detail, world understanding, and temporal consistency. Meta has acknowledged remaining gaps in audio-video synchronization and physically accurate fast motion, and the model is already restrictive on copyrighted content. First previewed in July alongside Muse Image, this is the first evidence it has moved into working beta.

Buzz. Block released it July 21, open source under Apache 2.0 and built on Nostr, with 15,600 GitHub stars in week one. Every human and agent holds their own cryptographic keypair rather than a shared bot token, so each agent has its own permissions and audit trail and can be revoked individually. Not ready to replace Slack: no mobile clients, no SOC 2, single-relay architecture despite the decentralized label, and Block-hosted relays are not end-to-end encrypted.

Stay building. 🤖