Perplexity chief Aravind Srinivas detailed Perplexity Computer, an AI system that forms a “team of agents” using as many as 20 models.
The platform’s agent harness coordinates models, tools and files while optimising privacy, performance and cost
Perplexity CEO Aravind Srinivas has outlined the company's vision for AI operating systems that can coordinate multiple artificial intelligence models, tools and files while balancing privacy, cost and performance.
Speaking during Intel CEO Lip-Bu Tan's keynote session at COMPUTEX 2026 on Tuesday, Srinivas described Perplexity Computer, launched earlier this year, as a system that can use up to 20 AI models and orchestrate tasks across different tools and files.
"It creates a team of agents, uses up to 20 different AI models, and it orchestrates across models, tools, and files in one single system," Srinivas said at the event organised by the Taiwan External Trade Development Council.
He said the system's "agent harness" is designed to coordinate different AI models while balancing intelligence, accuracy, privacy and cost through what he described as hybrid agentic inference. “This allows you to run smaller models locally," he said.
His remarks highlighted a growing focus within the AI industry on running some AI functions on personal devices while continuing to use data centres for more compute-intensive tasks.
"What we are showing today is just the start," the tech CEO said, adding that the future would involve more computing power in both data centres and local machines.
Describing the technology as a milestone for engineering on both the AI and chip sides, Srinivas thanked Intel for its partnership. "It's been really fun to partner with you and Intel on this," he said.
Earlier in the keynote, Tan said AI was reshaping how people use computing devices and highlighted on-device AI as a key area of focus.
"AI is profoundly impacting the way we use our devices. A major focus area for us is the use of AI on devices. Together with partners, we are at the forefront of advancing intelligence," Tan said.


























