v2.7.4: Your AI Employee Now Builds Slides, Invoices, and CVs That Actually Look Good
There is a particular kind of disappointment that comes from asking an AI to make you a presentation and getting back a wall of plain text pretending to be slides. You wanted something you could actually put in front of a client. You got an outline. This release closes that gap. Ask any employee for a presentation, an invoice, a CV, or a report, and what comes back is polished, light, and ready to send.
Presentations, the way you actually meant it
Ask an employee for a presentation, or slides, or a PowerPoint, and it builds one. No template hunting, no boilerplate. The first thing it does is ask which of two formats you want, because they serve different needs.
The first is a real, themed .pptx file, the kind you open and edit in PowerPoint or Office. It is built by a proper generator with ready layouts: title, section, bullets, two-column, big-number, native editable charts, quote, image. Light by default, professional out of the box, no hand-rolling.
The second is more fun. It is an online animated presentation, a self-contained file that opens in any browser and works fully offline. Elements morph and glide between slides, entrances stagger, charts are live, numbers count up. You can present it or edit it right there. And employees are now told to make these decks actually move, so you get real motion instead of a flat static wall.
If you pick the online version, the employee then asks whether you want it live at a URL, and if so, whether to put it behind a password or keep it public. You get back a link you can share or email. Pick a web address that is already taken and it quietly grabs the next free one instead of failing. You can even ask for a quick PNG of the first slide right in the chat, so you see the thing before you open it.
Documents that look styled, not dumped
The same care went into documents. Markdown and HTML now render into clean, properly formatted PDFs, real headings, code blocks, tables, and callouts, instead of the old flat text dump.
On top of that there are ready business templates. Ask for an invoice, a quote, or a CV and you get a professional, data-driven PDF. Choose a look from a set of themes, drop in your brand color and logo, and the employee remembers your style for next time, so the second invoice matches the first without you repeating yourself. Everything is light by default now, presentations and documents alike, so it all prints and shares cleanly.
Heavy PDF jobs used to be able to bog things down. Rendering now runs on a shared, queued service, so a burst of requests can no longer overload anything.
A real browser your employee can drive
The new playwright skill gives every employee a genuine headless browser running on the server. No extension, no local browser, nothing for you to set up. It can screenshot any URL full-page or just the viewport, scrape a page down to clean text or raw HTML, render a live web page to PDF, and run JavaScript in the page to read the result.
It runs through a safety queue, one job at a time with a timeout, so a heavy scrape can never swamp the machine. Worth being clear on the distinction: this is different from browser automation, which drives your own Chrome with your own logins. Playwright is for anonymous, server-side jobs where the employee just needs to go look at the open web and bring something back.
Its own SSH key, held properly
Ask an employee to connect to a server over SSH, or to hand over its key, and it now does the sensible thing. There is a new SSH skill, auto-installed for everyone. It reuses an existing key kept in a persistent folder rather than regenerating one, or creates a modern ed25519 keypair if it has none. It proposes a clear label for the key and asks you to confirm before generating anything. Its public key lives in its persona so it can hand it over instantly on request, and the private key never leaves.
See exactly which key paid for what
Voice runs on Gemini, and there is now an admin view that shows every voice interaction across the workspace at a glance: Nova, browser voice calls, and phone calls, with global totals and a breakdown per employee and per user. Each call shows the last four characters of the key used, the call type, duration, and tokens, so you can tell precisely who consumed which key and reconcile it against your Google billing. Expand any person and you get their full per-call history. If you run multiple keys, this is the difference between guessing and knowing.
Phone calls get Nova's conversation choice
Gemini phone calls now offer the same two conversation modes Nova has. One-at-a-time is a walkie-talkie rhythm where the assistant finishes and then it is your turn, and background noise will not cut it off, ideal for loud places. Interrupt-anytime lets you talk over the assistant and it stops to listen, smoothest in a quiet spot. You set it per number before the call.
Under the hood
Two more things worth knowing. Platform features moved off Google's soon-to-retire Gemini 2.5 line onto the current 3.x models, with each target checked against the live model list rather than blindly bumped, and web search keeps a free-tier fallback so no-billing keys still get answers. And a couple of chat fixes landed: switching your effort or thinking level in Claude chat mode now applies immediately instead of hanging, and Nova stays live cleanly instead of ever slipping into a reconnect loop.
The through-line for this release is simple. The work your employees produce should look like it came from a capable colleague, not a text box. Slides that move, documents you would actually send, a browser they can drive, and clear visibility into what everything costs.
Want to test the most advanced AI employees? Try it here: https://Geta.Team
