What happened
Perplexity’s announcement says Multiagent is a new effort level above Ultra for Computer’s most complex work. Computer is the company’s agent platform for researching complex questions and carrying out multistep work, including data analysis and app building. Perplexity says it introduced effort controls in September, offering four settings that let people choose how much effort Computer puts into a task. In those existing settings, one lead AI agent plans the task and delegates pieces to supporting agents.
Multiagent changes how that delegated work is coordinated. Perplexity says the lead runs a team of agents on different frontier models in parallel. The agents share findings directly with one another while they work, use evidence found by a teammate, question assumptions, and flag discrepancies for further investigation. The stated aim is to inspect a complex problem from multiple angles before the lead combines the work into one recommendation or requested output.
The announced default configuration names Claude Fable 5.2 as the planner and task divider. Supporting agents run Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon. Perplexity says users can name the models they want in a query. Fable 5.2 and GPT 6 Astra can join as supporting agents when a user requests them or when a particular part of the work calls for them. With Multiagent selected in Computer’s effort slider, the lead can assign up to 16 agents while also handling part of the work itself.
The agents can coordinate edits to shared files, according to the announcement. Perplexity gives the example of one agent updating revenue assumptions in a financial model and notifying another agent to use those figures in a valuation analysis. The lead remains the user’s single conversational contact, collects the team’s work, and delivers the final response. Users can guide the task through that conversation without messaging every agent individually, see the agents and models in use as the task runs, and stop the full team by stopping the thread. For a follow-up task, the lead tells affected agents what changed and the team continues its existing work.
Perplexity positions Multiagent for assignments with separate but connected parts: financial forecasting across scenarios, annual-report review, deep analysis of competing explanations, deal data-room diligence, and preparation of questions and analysis for tax professionals. Its comparison distinguishes Ultra, where supporting agents report only to the lead; Multiagent, where agents exchange findings and work on shared deliverables; and Model Council, where several models independently investigate the same question before Computer combines their findings. Multiagent is available on the web to Perplexity Max subscribers, selected from Computer’s effort slider for a new task. Perplexity does not state a separate feature price in the announcement.
What the archive adds
New Runtime’s earlier coverage of Perplexity Teammates on 2026-11-06 described persistent work agents for Perplexity Max subscribers. Multiagent is a new event and a more specific coordination mode within Computer: the announcement describes a temporary team that shares findings and common deliverables during a task.
Trend signal
This supports harness architecture outlives model choice. Perplexity presents Fable, Opus, GPT, and Gemini as roles inside one managed workflow rather than as mutually exclusive products. The useful product boundary here is the planner, delegation structure, shared work, visibility, and stopping control. The announcement does not demonstrate that models can be swapped freely in every workflow, but it makes cross-model orchestration a first-class capability.
Why now
Perplexity places the release after its earlier effort controls and above Ultra in the effort hierarchy. Its own comparison clarifies the new choice: use Ultra for delegated exploration that reports back to one lead, Multiagent for interconnected work requiring collaboration, and Model Council for independent perspectives on one question.
Why it matters
The release turns a design distinction often left to custom agent harnesses into a selectable product mode. Builders should not treat “more agents” as a generic quality setting. Multiagent’s stated value depends on whether the assignment can be divided, whether dependencies between parts are important, and whether evidence should be checked by another worker before a final answer is assembled.
The practical mechanisms are more consequential than the number of agents. Direct finding-sharing can prevent repeated research; discrepancy flags create an explicit route for revisiting an assumption; shared-file coordination makes outputs interdependent rather than a bundle of isolated drafts. A common conversation also gives the user one control point while preserving task-level visibility into the models and agents involved.
For teams, this provides a concrete comparison framework. A bounded diligence or forecasting task may merit collaborative workers because one conclusion changes another worker’s input. A question seeking independent judgment may instead be better served by Model Council. An open-ended task with parallel avenues but fewer dependencies may fit Ultra. The relevant evaluation is whether the coordination improves a reviewed deliverable enough to justify its extra work, not whether it activates the largest team.
What happens next
- Perplexity will expand the model and role controls exposed in Multiagent. (6 months, 60%). Watch for: named team templates, per-agent instructions, or persistent model preferences.
- Users will ask for measurement of collaborative quality against Ultra and Model Council. (3 months, 70%). Watch for: published completion, error, cost, or review-time comparisons.
Limits and uncertainty
Perplexity describes the product design and intended applications, but publishes no benchmark, customer result, completion-time comparison, error rate, or cost per completed task. It does not quantify how often agents catch one another’s errors, how disagreements are resolved, or how concurrent edits to shared files are controlled. The announcement confirms web access for Max subscribers but does not state usage quotas, consumption rules, task-duration limits, or whether model choices alter those limits. The operational value therefore remains unverified outside the company’s examples.
Skill to build now
Take one existing agent workflow and classify it before adding workers. Use a lead-only delegation pattern when subtasks can return independently. Use a collaborative pattern only when one worker’s evidence or file edit changes another worker’s output; define the shared artifact, the handoff condition, and the person or agent responsible for resolving a conflict. For an independent judgment call, run parallel investigations without sharing intermediate conclusions, then compare them afterward.
Reader takeaway
Perplexity’s Multiagent is not simply a larger model ensemble. It is a Computer mode for connected assignments in which agents share evidence, challenge assumptions, and coordinate shared work under one lead. Choose it when those dependencies are real, and validate it against the review time and output quality of simpler alternatives.