Perplexity Computer Launches Multiagent: 16 AI Agents Cross-Validate Each Other

Perplexity adds a Multiagent tier to Computer, where a lead agent orchestrates up to 16 heterogeneous sub-agents for cross-validation.
Perplexity has added a Multiagent mode to its Computer product line, pushing the "effort level" one tier above Ultra. In this mode, a lead agent orchestrates up to 16 sub-agents running on Claude, GPT, and Gemini, which share intermediate findings and review each other's work through two core mechanisms: shared findings and cross-validation. The article argues this turns "how much compute are you willing to spend" into an explicit product dimension, trading time and compute for more reliable multi-perspective conclusions—suitable for research, due diligence, and complex code review, but overkill for everyday Q&A. Benefits come from parallel decomposition, shared findings, and heterogeneous validation, while costs include extra token and time overhead, orchestration-layer bottlenecks, and the lead agent's decomposition quality setting the overall ceiling.
Perplexity has added a Multiagent mode to its Computer product line, pushing "effort level" one tier above Ultra. According to the official description, this mode features a lead agent that orchestrates up to 16 sub-agents, each running on Claude, GPT, and Gemini models respectively. The sub-agents share intermediate findings with one another and review each other's work.

What Multiagent Actually Does
Traditional AI assistants work in a single-threaded fashion: one model takes a task, reasons through it from start to finish, and only it is "checking" itself along the way. Multiagent changes this structure—the lead agent is responsible for breaking down the task, distributing subtasks, and then aggregating the outputs of up to 16 agents.
The key lies in two mechanisms: first, shared findings—sub-agents don't work in isolation; intermediate results are visible to other agents; second, cross-validation—each agent's output is reviewed by its peers. This is essentially bringing the human team practice of "pair review" down to the model layer.
One detail worth noting: these sub-agents don't all run on the same model. Claude, GPT, and Gemini are deployed simultaneously, meaning tasks can benefit from the different preferences and capabilities of each model, and the validation step naturally gains heterogeneity—different models' judgments on the same piece of work tend to be more discriminating than a model reviewing itself.
Lead Agent is the orchestrator in a multi-agent system. Its job is not to directly complete the user's task, but to decompose it into subtasks that can be executed in parallel or sequence, assign them to appropriate sub-agents, and finally merge the various outputs and resolve conflicts. The quality of its work is often more critical than that of any individual sub-agent—if the orchestration layer decomposes the task poorly, no amount of downstream parallel compute will help, as it will be wasted on the wrong division of labor. This is also the most fundamental difference between multi-agent architecture and the traditional "one model from start to finish" approach: the latter only has reasoning, while the former adds planning and coordination.
Heterogeneous model validation refers to having models from different vendors with different training data review each other's outputs. Same-model self-review tends to fall into the trap of "self-consistent but actually wrong," because it will re-walk the same path using its own reasoning preferences; whereas Claude, GPT, and Gemini differ in training corpus, alignment strategies, and reasoning styles, and often reach different judgments on the same intermediate result, thereby exposing individual models' hallucinations or logical leaps. This heterogeneity is the precondition for multi-agent cross-validation to work, and explains why Perplexity deploys all three models simultaneously rather than simply stacking the same model.
"Effort Level" Becomes a New Product Dimension
Perplexity positions Multiagent as "an effort level above Ultra," and this wording itself is noteworthy. In the past, differentiation in this product category revolved mainly around model selection, context length, and retrieval quality; now there is an explicit "how much compute are you willing to spend" dimension.
Users are effectively making a cost-versus-depth tradeoff: the standard tier prioritizes response speed, the Ultra tier pursues more thorough reasoning, and the Multiagent tier sacrifices time and compute in exchange for more reliable multi-perspective conclusions. For tasks with low tolerance for error—research, due diligence, complex code review—this trade is worthwhile; for everyday Q&A, it is clearly overkill.
This design also exposes a reality of current AI products: no matter how strong a single model is, it can produce self-consistent but actually wrong reasoning in long-chain tasks. Introducing multiple heterogeneous models to catch each other's errors is one of the more engineering-feasible mitigation approaches available today.
Test-time Compute refers to the scale of compute invested during the answer-generation phase, as opposed to parameter scale during training. Recent research and practice have shown that having a model think longer, sample multiple times, or self-verify before answering can significantly improve accuracy on complex tasks. By making "effort level" an explicit tier, Perplexity is essentially turning test-time compute from a behind-the-scenes parameter into a user-perceivable, selectable cost item—users trade latency and quota for more thorough reasoning and validation, which is exactly the product logic behind the phrase "Multiagent above Ultra."
Benefits and Costs of Multi-Agent Architecture
From an architectural standpoint, Multiagent's benefits come from three directions: parallel decomposition allows sub-tasks to be more focused; shared findings reduce redundant exploration; cross-validation lowers the risk of individual model hallucination. Theoretically, the combined coverage of 16 agents is far greater than a single agent's one-pass reasoning.
The costs are equally obvious. Communication, aggregation, and mutual review among agents all consume extra tokens and time, and the orchestration layer itself can become a bottleneck or source of error. More critically, the lead agent's task decomposition quality determines the overall ceiling—if the decomposition itself is off, more sub-agents will only mean division of labor in the wrong direction.
Additionally, cross-model collaboration is not friction-free in engineering terms. Different models differ in output format, reasoning style, and context handling, and the aggregation step requires extra alignment work. The stability of this mechanism in real complex tasks still needs more practical use to verify.
What It Means for Developers and Power Users
For developers, the signal from Multiagent is that orchestration capability is becoming part of product competitiveness. Models themselves are increasingly like interchangeable components, and "how to organize multiple models to work together" is becoming a new differentiation space. This also aligns with the recent popularity of multi-agent frameworks.
For power users, the actual value of this tier depends on the task type. Scenarios that require multi-source comparison, cross-verification of conclusions, or simultaneous coverage of multiple tech stacks or explanation paths are Multiagent's home turf; for single-point fact queries, simple rewrites, or format conversions, this tier is not economical.
Perplexity has not yet disclosed in public information the specific latency and quota differences between tiers, and the actual experience will need to be judged by in-product performance. What is certain is that making "multi-agent cross-validation" a user-selectable tier is itself a meaningful attempt—it hands the choice of compute allocation back to the user.
Multi-agent orchestration frameworks are a development paradigm that has emerged in recent years, with representatives including AutoGen, LangGraph, and CrewAI. The core idea is to organize multiple agents with different roles or models through message passing, task graphs, or review loops to collaboratively complete complex tasks. Perplexity's Multiagent can be seen as a product-level implementation of this paradigm: models are gradually being treated as interchangeable components, and the real competitiveness shifts to "how to organize multiple models to divide labor, share context, and cross-validate." For developers, this means that orchestration-layer design capability, state management, and conflict resolution strategies are becoming technical issues as important as model selection.
FAQ
Which models does Multiagent use? According to Perplexity's explanation, the sub-agents run on Claude, GPT, and Gemini, orchestrated by the lead agent.
What is the maximum number of agents? Up to 16.
What is its relationship to the Ultra tier? Multiagent is an effort level above Ultra, aimed at tasks requiring deeper reasoning and cross-validation.
How do the agents collaborate? Sub-agents share intermediate findings and review each other's work output.
Related articles

Programmers in the AI Era: How Deep Should You Learn Legacy Tech?
How should programmers transform under AI's impact? How deep should legacy tech be learned? Drawing on a Bilibili creator's share, this breaks down the levels of embracing AI, the shift from process-driven to project-driven learning, and how lower technical barriers affect junior and mid-level roles.

OpenAI Releases GPT-6.2 Dual Models Astra and Sol, API Price Cut by 20%
OpenAI releases GPT-6.2 Astra and Sol, with Astra's API price cut 20% to $8/$40 per million tokens and ranking second on the Artificial Analysis Intelligence Index. This article analyzes the dual-model strategy, price-cut impact, and developer selection advice.

Amazon and Google to Jointly Hold Nearly 30% of Anthropic
After Anthropic's IPO, Amazon will hold just under 15% with an option to increase, and Google nearly 14%, together controlling close to 30%. This article analyzes the strategic logic, governance challenges, and impact on AI competition.