What changed

Perplexity added Multiagent to Computer, its platform for multistep research, as a new effort level. It sits above the Light, Standard, High and Ultra settings introduced in September 2026. The company did not publish any accuracy or speed benchmarks for the feature.

How it works

A lead agent plans the task, splits it into parts and can assign up to 16 supporting agents, while also doing part of the work itself. In Perplexity's example, Claude Fable 5.2 plans the task and distributes work to agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. Those three models form the default team.

  • Direct sharing: Agents use evidence found by others, question assumptions and flag discrepancies.
  • Shared files: Multiple agents can coordinate edits to the same file.
  • Single point of contact: Users talk only to the lead, which merges findings into one response.
  • Control: Users can see which agents and models are running, and stopping the thread stops the whole team.

How it differs from Ultra and Model Council

Perplexity now offers three ways to put several models on one problem. In Ultra, supporting agents report only to the lead. In Multiagent, agents talk to each other and edit shared deliverables. Model Council has several models answer the same question independently and shows where they agree and disagree.

Who it is for

Perplexity positions Multiagent for large, interconnected work that can run in parallel, citing examples such as deal data room diligence, multi-scenario forecasts, annual report review and year-end tax preparation. The feature is currently web only and limited to the Perplexity Max plan, with no local or quantized option.