Perplexity’s Multiagent lets rival AI models share work under one lead
Perplexity is adding direct collaboration between frontier models to Computer, letting a lead agent coordinate up to 16 supporting agents. Company-reported benchmarks suggest mixed teams can outperform repeated copies of one model, but leave customer costs and production reliability unresolved.
- Multiagent lets agents exchange findings and coordinate shared files while one lead assembles the result.
- The new effort level is available to Perplexity Max subscribers on the web.
- Reported benchmark savings compare mixed teams with repeated copies of one model; they do not establish cheaper everyday use.
Perplexity, an AI search company, is giving its Computer agent platform a new way to divide complex assignments: a lead AI can coordinate up to 16 supporting agents using different frontier models, with those agents sharing evidence and checking one another’s work. The company announced the Multiagent rollout on November 17, extending a platform that already coordinated specialized models for multistep tasks.
Computer handles projects such as building websites, drafting reports and analyzing datasets by delegating work to software agents—AI models that use tools to carry out tasks. In its February launch update, Perplexity said Computer could already orchestrate 19 models in parallel, connect to hundreds of services and run in the cloud. Multiagent adds a collaboration mode to that existing system.
The consequential change is how information moves. Ultra and Multiagent use the same lead model, but under Computer’s existing Ultra setting, supporting agents report only to the lead. In Multiagent, they can question one another’s assumptions and connect findings before the lead delivers a single answer, Perplexity says in its announcement. The feature’s proposition is coordination across interconnected work.
Different models, shared deliverables
Perplexity says Claude Fable 5.2, an Anthropic model, plans the assignment and divides it among a default supporting team running Anthropic’s Claude Opus 5.5, OpenAI’s GPT 6.1 Sol and Google’s Gemini 4 Carbon. Users can request particular models; Fable 5.2 and OpenAI’s GPT 6 Astra become supporting agents only when requested or when a task component calls for them. Perplexity controls the default team and coordination, while the underlying models come from those competing developers.
The company’s examples explain where collaboration could matter. One agent might change revenue assumptions in a shared financial model, then tell another to incorporate those figures into a valuation. In deal diligence, agents could connect a contract provision with its implications for revenue while reviewing financial and operating records separately.
Other proposed uses include comparing financial forecasts, testing competing research explanations, reviewing annual-report disclosures and preparing tax-planning analysis. These are proposed workflows, rather than demonstrated professional outcomes. Perplexity frames flagged accounting discrepancies as material for an auditor’s review and tax analysis as preparation for a tax professional.

