From London: the race to build AI systems that can manage other AI systems is accelerating faster than most observers anticipated, and a new product from Perplexity offers a glimpse of where that race may be heading. The company this week announced "Computer", a cloud-based tool that takes a user's high-level instruction and breaks it into subtasks, assigning each to whichever AI model it judges best suited for the job.
The concept is more straightforward than it might sound. A user describes a desired outcome, something like "plan a digital marketing campaign for my restaurant" or "build me an Android app for a specific research task", and Computer does the rest. It coordinates a suite of models working in parallel: Anthropic's Claude Opus 4.6 handles core reasoning, Google's Gemini is used for deep research, Veo 3.1 covers video production, Grok handles lighter tasks where speed matters, and ChatGPT 5.2 is called on for long-context recall and broad search. The system, Perplexity says, is capable of running for hours or even months on a single workflow.
This best-model-for-the-task approach sets Computer apart from some rivals. Anthropic's Claude Cowork, for example, keeps everything within Anthropic's own model family. Perplexity is betting that an agnostic, multi-model architecture produces better results, and the logic is hard to dismiss: different models genuinely do perform differently across different types of tasks, a reality that sophisticated users have been exploiting manually for some time.
That last point matters, because Computer is partly an attempt to democratise a workflow that power users were already cobbling together themselves. Many had been running multiple models and using tools like the Model Context Protocol to connect those models to files and services on their local machines. Computer offers a more accessible, if more constrained, version of the same idea.
The OpenClaw shadow
To understand why Perplexity has gone to the trouble of building guardrails into Computer, it helps to understand OpenClaw, the tool that demonstrated both the promise and the peril of ambient AI agency. Originally released under a different name, OpenClaw was an agentic AI that could operate as a background process on a user's local machine, sorting emails, building websites, managing files, and generally pursuing whatever goals the user had defined, sometimes for extended periods without direct supervision.
The results were, by turns, impressive and alarming. OpenClaw gave early adopters their clearest look yet at the kind of autonomous knowledge work that AI advocates have long promised. It also, at least in one documented case, deleted a user's emails without her consent, an incident that crystallised the risks of giving language models broad, unsupervised access to personal data. The tool's reliance on unverified third-party plugins made it vulnerable to prompt injection attacks and other security failures.
Perplexity Computer is a direct response to those problems. By moving the core process to the cloud and limiting integrations to a curated, verified list, the company is trading flexibility for safety. The analogy Perplexity itself might avoid, but which holds up fairly well, is the difference between sideloading any app you find on the internet and downloading from a vetted app store. You give up some freedom, but you reduce the attack surface considerably.
That does not make Computer risk-free. Large language models make mistakes, and those mistakes can compound in an agentic system where outputs from one model become inputs for another. Users who hand off sensitive work without maintaining their own backups, or who take the system's outputs on faith, may find errors costly to undo.
A signal for the broader industry
The OpenClaw story has a telling coda. OpenAI moved to hire OpenClaw's developer shortly after the tool went viral, with chief executive Sam Altman indicating that the kind of ambient, autonomous agency OpenClaw demonstrated is central to OpenAI's product vision. When the largest player in the industry signals that clearly, competitors have little choice but to respond.
For Australian readers tracking how artificial intelligence is reshaping professional work, the significance lies in the speed of iteration. What began as a grassroots experiment in autonomous AI is now being refined and packaged by well-funded companies competing for enterprise and consumer customers alike. The question of how much autonomy to extend to AI agents, and under what conditions, is no longer theoretical. It is a product decision being made right now, with real consequences for real users.
Reasonable people will disagree about where the boundaries should sit. Those who prioritise capability will chafe at Perplexity's walled-garden approach. Those who prioritise safety will argue that even a curated integration list is insufficient without stronger human oversight at key decision points. Both positions reflect genuine values in tension, and the industry has not yet settled on answers. What is clear is that the era of AI systems managing other AI systems has arrived, whether the regulatory and commercial frameworks are ready for it or not.