Perplexity has added Multiagent, a fifth effort level for Computer, its agent platform for research and multistep work, and it is the first setting where the supporting agents talk to each other instead of only to the boss. Perplexity announced it in a blog post and is limiting it to Max subscribers on the web. The design bets that models from different labs catch each other’s mistakes better than one model checking itself.

Computer already worked by delegation. Since September, users have picked from four effort settings, and at each one a lead agent plans the job and hands parts of it to helpers. Multiagent sits above the top of those, called Ultra. The lead now runs a team on different frontier models, and the team works in parallel.

The default roster mixes three vendors. Claude Fable 5.2 plans the task and splits it among agents running Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon. Fable 5.2 and GPT 6 Astra can also join as supporting agents, but only when the user asks for them or when a piece of the task calls for them. Users can name the models they want in the query.

What separates this from Ultra is the traffic between agents. In Ultra, helpers answer to the lead alone. In Multiagent, an agent can build on evidence a teammate found, push back on a premise, or raise an inconsistency for a closer look. Agents also coordinate edits to shared files. Perplexity’s example: one agent revises the revenue inputs of a financial model, then tells a second agent to feed those numbers into its valuation.

The cap is 16 agents per team, and the lead keeps a share of the work for itself. It stays the single contact point, so users steer through one conversation and never message individual agents. On a follow-up, the lead briefs only the agents touched by the change and the rest carry on. The interface shows each agent and the model it runs, and stopping the thread stops everyone.

Perplexity points the feature at jobs with separable parts that need cross-checking: financial forecasts with several scenarios, annual report review, deal data room diligence, tricky tax returns with year-end planning, and deep analysis of competing explanations. Several of those are work where a wrong figure is expensive. The company frames the output as material for an auditor or tax professional to review, not a replacement for one.

Perplexity also draws a line between its three multi-model options. Ultra suits questions with no fixed method, where several approaches get tried. Model Council sends multiple models at one question separately and then maps their agreement and disagreement. Multiagent covers big assignments whose parts depend on each other. The split matters for anyone choosing a setting: independence is the point of Council, while collaboration is the point of Multiagent.

The announcement is all design and no measurement. Perplexity does not publish benchmark results, error rates, or a comparison of Multiagent against Ultra on the same tasks. It also does not say how much compute a run consumes or how long a 16-agent job takes. Agents that read each other’s output can reinforce a shared error as easily as they catch one, and the post offers no evidence on which tendency wins.

Finance and diligence teams that already use Computer should run one past engagement through Multiagent and Ultra side by side, then compare the discrepancies each flags against what the human reviewer found.

Source: Perplexity, “Introducing Computer’s Multiagent effort level,” Perplexity blog.