Perplexity just rolled out a new AI called Computer that hands off jobs to a team of other AI agents.
This setup lets people give a big goal, then watch as the main agent breaks it down and gets different specialized AIs to handle the pieces using whichever model fits best.
Right now, only folks with a Perplexity Max plan can try Computer, which the team describes as something that builds full step-by-step plans and keeps working on them anywhere from hours up to months at a stretch.
You simply tell it what end result you want, for instance, mapping out and running a nearby online promo push for an eatery you own, or putting together a mobile app on Android designed for a certain type of work research you do.
After that, the system thinks up smaller steps on its own and sends each one to the right helper agent powered by the model Perplexity picks as strongest for that job.
At its heart, the thinking part relies on Anthropic’s latest Claude Opus 4.6 version, while heavy digging pulls from Gemini, pictures come via Nano Banana, videos get made with Veo 3.1, quick, easy chores go to Grok for faster replies, and big memory or broad lookup jobs tap ChatGPT 5.2.
Picking the top tool per job sets this apart from rivals such as Claude Cowork that stick strictly to one family of models from Anthropic.
Everything takes place up in the cloud and comes with ready-made connections already built in.
Perplexity explains that individual jobs happen inside separate safe computing spaces equipped with genuine file storage, a working web browser, and actual useful connections to services.
Part of the thinking behind it stems from how advanced users have already been juggling several models on their own, matching each to what it does well and linking them to personal files or apps through things like the Model Context Protocol setup.
Computer shifts that idea into a simpler package so regular people can get similar results without wrestling with complicated configurations.
Even so, the basic aim stays identical, letting smart agents use carefully chosen models to tackle work that touches your documents, online accounts, and programs.
Leading up to this, you had OpenClaw sitting as the closest earlier example of the same kind of thinking.
For anyone who missed the big buzz around OpenClaw here is a fast recap—it started life under names like ClawdBot and later Moltbot before settling on OpenClaw as an independent AI helper that ran quietly in the background on personal computers using powerful language models to take on all sorts of jobs from digging through old messages to whipping up entire sites or pretty much anything users dreamed up.
With enough access rights and the correct add-ons, it gained the power to read, change, or wipe out files on the device, going way past limits most chat-style AIs could reach even with Model Context Protocol links.
People fed it ongoing direction through special documents named things like USER.MD MEMORY.MD SOUL.MD or HEARTBEAT.MD that spelled out its purpose and let it keep pushing forward on its own, often for extended periods with little human nudging.
That freedom delivered some eye-catching wins, giving early tastes of the deep thinking and real-world output that supporters of agent-style AI have long promised.
At the same time, though it ran into major slip-ups, got hit hard by sneaky prompt tricks, and faced other safety gaps largely because of the loose, uncontrolled world of community-made add-ons.
The very same flexible setup that once spun up a trendy fake Reddit full of bot characters also, in at least one reported situation, wiped out somebody’s inbox emails without permission.
Perplexity Computer steps in, trying to fix several of those weak points. For starters, the main action stays safely in the cloud instead of sitting directly on your own hardware.
Next, it operates inside a tightly controlled space featuring only hand-picked approved connections rather than the anything-goes style OpenClaw allowed.
Think of the comparison this way, even if not perfect you might picture OpenClaw as the free-for-all early internet, while Computer feels more like the managed world of an official app marketplace from Apple.
Sure, that means fewer wild possibilities, but it also keeps away random untrusted code that could poke around dangerously in your setup.
Risks have not vanished completely, of course. Large language models still mess up sometimes, and those errors might cause real trouble, especially when dealing with important unsaved information or when users skip double-checking what comes back.
In short, Perplexity Computer works to tame, polish, and safely box in the raw, exciting energy that made the OpenClaw trend explode while going head-to-head against tools like Claude Cowork through smart choices of the right model for every little step.
Plenty more companies in the AI space will likely jump on similar ideas soon enough.
After all, OpenAI brought aboard the person who built OpenClaw, and its leader, Sam Altman, has hinted that pieces of what that project showed off will play a key role in where the company heads next.
Other Stories You May Like