Market · AI in practice

Perplexity puts agents on different models to check each other's work

Multiagent, a new effort level in Perplexity's Computer, splits a task across up to 16 agents on models from different labs, which share findings and check each other's work. The examples come from finance.

This note is in the summary for the week of 16 NOV 2026 to 22 NOV 2026 .

Company or topicPerplexity

What the sources report

Computer, Perplexity's agent platform, has a new effort level, Multiagent, above Ultra. In Ultra, a lead agent plans the task and delegates parts to supporting agents, which report only to it. In Multiagent, the supporting agents talk to each other: one uses evidence another found, questions an assumption or flags a discrepancy for more investigation. At the end, the lead brings it all into a single response.

In the default team, which mixes labs, Claude Fable 5.2 plans and divides the task among agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon, and users can ask for other models in the query itself. The lead can assign up to 16 agents, which coordinate edits to shared files. In Perplexity's example, one agent updates the revenue assumptions in a financial model and tells another to use those figures in the valuation.

Perplexity lists use cases from a finance team's work: forecast scenarios built in parallel and compared, annual report reviews with footnotes cross-checked to flag inconsistent figures for the auditor, data room diligence across contracts and financial records, and year-end tax planning to bring to a tax professional. The announcement gives no figure for accuracy gains.

My take

Cross-checking across models from different labs is a practical way to catch a wrong assumption before it reaches a forecast or a diligence report. In work with lots of spreadsheets and documents, the costly mistake is usually in the assumption nobody checked.

Author

Adriano Meirinho · Co-founder & CAIO (Chief AI Officer) at Celcoin. About

All notes