Perplexity 为 Computer 平台推出 Multiagent 模式
Perplexity Launches Multiagent to Coordinate Frontier AI Models
Perplexity 于 2026 年 11 月 17 日推出 Multiagent,这是 Computer 智能体平台新增的 effort 设置,由 lead 智能体把任务分派给最多 16 个支持智能体,跨多个前沿模型分析复杂任务、共享证据并交叉核对后再给出单一回复。该模式位于 Ultra 之上,支持智能体之间直接交换发现、质疑假设并标记分歧,而 Ultra 的支持智能体只向 lead 汇报。默认支持团队由运行 Claude Opus 5.5、GPT 6.1 Sol 和 Gemini 4 Carbon 的智能体组成,Claude Fable 5.2 担任 lead,Fable 5.2 与 GPT 6 Astra 仅在用户要求或任务需要时加入。
正文
Perplexity on November 17, 2026 introduced Multiagent, a new effort setting for its Computer agent platform whose lead can assign up to 16 supporting agents across several frontier models to analyze complex tasks, share evidence and cross-check their work before delivering a single response.
The new level sits above Ultra and changes how information moves inside Computer’s agent teams. Perplexity says that, unlike Ultra’s supporting agents, which report only to the lead, Multiagent’s agents can exchange findings directly, question assumptions and flag discrepancies for further investigation. The design is intended to identify conflicting assumptions before the lead combines the team’s work into the requested recommendation or other deliverable.
Several Models on One Assignment
Claude Fable 5.2 serves as the lead, planning the work and dividing it among agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. Those three models form the default supporting team. Fable 5.2 and GPT 6 Astra can also join as supporting agents, but only when the user requests them or a particular task component calls for them.
Perplexity is aiming the mode at assignments with distinct but connected parts. In an annual-report review, agents can examine financial statements, footnotes and management commentary separately, then compare their findings to flag inconsistent figures or disclosures for an auditor. For deal data-room diligence, separate agents can review financial records, contracts and operating documents in parallel and connect details such as a customer contract that changes the revenue outlook.
Shared Files and User Controls
Agents can also coordinate edits to shared files. Perplexity gives the example of one agent updating revenue assumptions in a financial model and notifying another to use those figures in its valuation analysis. On a follow-up, the lead tells affected agents what changed and the team continues its existing work; stopping the thread stops the entire team. The lead can also take on part of the assignment itself.
Where Multiagent Fits in Computer
Multiagent extends the effort controls introduced in September 2026, which offered Light, Standard, High and Ultra settings. Perplexity selected the coordinating model and reasoning level for each setting, while custom controls let users choose a specific model and reasoning level themselves. Lower settings used less expensive models, and credit usage depended on the work Computer performed.
Perplexity positions Model Council for judgment calls that benefit from independent perspectives: several models investigate the same question separately, after which Computer combines their findings and shows where they agree and disagree. Ultra is intended for open-ended problems that require several approaches to be tested, while Multiagent is intended for large assignments with interconnected parts that can proceed in parallel.
The annual-report example shows the intended division of responsibility: agents compare figures and disclosures across documents, while an auditor reviews the inconsistencies they flag. Multiagent’s coordination is therefore framed as preparation for professional scrutiny, with one lead still responsible for assembling the final response.
AI 热点