AI News
AI News AgentProductPerplexity3 min read

Perplexity launches Multiagent for complex tasks

Perplexity is launching Multiagent, a new Computer level that coordinates up to 16 AI agents to analyze complex tasks in parallel. The agents share findings, review assumptions, and work on shared deliverables, and the feature is available to Max subscribers on the web.

Perplexity is adding a new level of effort to Computer: Multiagent coordinates up to 16 AI agents working in parallel, sharing findings and reviewing one another’s work before delivering a joint answer.

The feature is designed for assignments with several connected parts, such as analyzing a company, reviewing financial documents, or building a forecast with different scenarios. It is not just about requesting several answers: agents can use their colleagues’ discoveries, challenge assumptions, and flag discrepancies for further investigation.

How Multiagent works

Computer uses a lead agent to organize the work. That agent breaks down the task, assigns parts to other models, and keeps the conversation with you. You do not have to talk to each agent separately.

The default team combines Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon. The lead agent, based on Claude Fable 5.2, decides how to divide up the assignment and then brings the findings together in a single answer with the requested results.

The agents can also coordinate changes to shared files. For example, one agent can update a financial model’s revenue forecasts and notify another agent to use those figures in its valuation analysis.

You can see which agents are working and which models each one is using as the task progresses. If you stop the conversation, the entire team stops. In a follow-up task, the lead agent informs the affected members about the changes, and the work continues from where it left off.

What it is for

Multiagent is designed for work where the different parts can be analyzed at the same time but then need to be connected. Perplexity highlights several use cases:

  • Financial forecasts: create scenarios with different assumptions and check how the results change when revenue, costs, or market conditions vary.
  • Annual report reviews: examine the financial statements, notes, and management commentary separately to find figures or statements that do not line up.
  • Deep research: compare possible explanations for a business or research problem and check which conclusions are supported by the available sources.
  • Transaction due diligence: review contracts, financial records, and operational documents, and connect findings that could affect the business valuation.
  • Taxes and year-end close: organize income and deductions, research tax rules, and compare scenarios to prepare questions for a professional.

Multiagent, Ultra, or Model Council

Perplexity places Multiagent above Ultra in Computer’s effort control, but each option organizes the work differently.

  • Ultra: a lead agent delegates parts of the assignment and receives results from supporting agents without those agents collaborating directly with one another.
  • Multiagent: agents share findings, review one another’s work, and contribute to shared deliverables.
  • Model Council: several models study the same question independently, and Computer shows where they agree and where they disagree.

The practical difference matters. Multiagent is a better fit when a large task has related parts that can move forward in parallel. Model Council is more useful when you want independent perspectives on the same decision. Ultra is suited to open-ended problems that require testing several approaches but do not necessarily need direct collaboration between agents.

Multiagent is available to all Perplexity Max subscribers on the web. To use it, select the option in Computer’s effort control when starting a new task.

The change points to a more coordinated way to use AI agents: not just generating answers, but dividing up a project, cross-checking its parts, and bringing them back together. Even so, in finance, taxes, or business decisions, the result still needs human review, especially when a discrepancy could have real-world consequences.