Perplexity's Computer can now run up to 16 AI agents that check each other
Perplexity added a Multiagent mode where a team of AI agents from rival labs shares findings and flags each other's errors before you see one answer.
The short version
- Multiagent is a new effort level in Perplexity Computer, above the existing Ultra setting.
- A lead agent can assign up to 16 agents, and it does part of the work itself.
- The default team uses Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon.
- Available to Perplexity Max subscribers on the web only; Max is listed at $167 per month billed annually.
- Perplexity published no accuracy or speed numbers, and no credit cost per task.
Why it matters
Wrong answers from a single chatbot are a daily cost for anyone who uses AI for research or analysis. Perplexity is betting that models from different companies, checking each other, will catch more of those errors. With no published benchmarks, that is still a claim and not a result. The habit behind it is one you can copy for free.
What Perplexity added
Perplexity has added a mode called Multiagent to Perplexity Computer, its platform for research and multistep work, as MarkTechPost reported. An agent here is an AI that takes steps on a task instead of only answering a question. Multiagent sits at the top of Computer's effort settings, above Ultra.
Those effort settings arrived on September 17, 2026, and came in four levels: Light, Standard, High and Ultra. Each level chooses which model acts as the orchestrator, the one that plans and directs the work, and how deeply it reasons. The lower levels use cheaper models.
Multiagent changes who talks to whom. In Ultra, helper agents report only to the lead. In Multiagent, they also talk to each other.
How the agent team works
The lead agent plans the job, hands parts to supporting agents, and does some of the work itself. It can assign up to 16 agents. In Perplexity's own example, Claude Fable 5.2 is the lead. It splits the work among agents running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon, which form the default team. Fable 5.2 and GPT 6 Astra can join as supporting agents only when requested or needed.
The agents work in parallel. They use evidence that others found, question each other's assumptions and flag discrepancies. They also coordinate edits to shared files. In one example, one agent updates revenue assumptions and tells another to reuse them in a valuation.
You deal only with the lead, which merges everything into one response. If you send a follow-up, the lead tells the affected agents what changed, and the team carries on. You can see which agents and models are running, and you can name particular models in your query. Stopping the thread stops the whole team.
Three ways to put models on one problem
Perplexity now has three ways to bring several models to one question, and they differ in how much the models interact.
- Ultra: one lead delegates to helpers, and the helpers report only to the lead.
- Multiagent: the same lead model coordinates helpers that talk to each other and edit shared deliverables.
- Model Council: several models answer the same question independently, and Computer shows where they agree and disagree. It came to Computer in July 2026.
Perplexity points Multiagent at large, interconnected jobs that can be split up and run in parallel. Its examples are deal data room diligence (combing through the documents in a business sale), multi-scenario forecasts, annual report review, research that tests competing hypotheses, and year-end tax preparation.
MarkTechPost also compares it with rivals. Perplexity's distinguishing feature is mixing models from Anthropic, OpenAI and Google in one team. Kimi and Grok use their own model families, and Kimi goes for scale, with up to 300 sub-agents, rather than depth of debate.
Who can use it, and what is missing
Multiagent is open to all Perplexity Max subscribers and runs on the web only. You pick it with the effort slider when you start a new task. It runs in Perplexity's cloud, with no local option. Perplexity's pricing page lists Max at $167 per month billed annually, and the plan includes 10,000 Computer credits a month.
Several things are not known. The launch post does not say how many credits a Multiagent task costs, so it is hard to judge how far the monthly allowance goes. The context length, meaning how much material the system can hold in view at once, is not disclosed. Perplexity also published no accuracy or speed numbers.
That last gap matters most. The case for Multiagent is that cross-checking catches mistakes, but the launch gives no measurement of how many. Treat it as a design choice that sounds sensible, not a proven improvement.
The idea you can borrow today
Think of it as a small office of reviewers, where each one checks the others before anything reaches you. The value comes from the reviewers being different, which is why Perplexity mixes models from three companies. Models from one family may repeat the same blind spots.
You can use that habit without paying for Max. Take an answer from one chatbot and ask a second, different chatbot to find errors, weak assumptions and claims that need a source. Then check the disputed points yourself. It is slower than a built-in team, but it works on any task where a wrong answer would cost you time.
What to watch
- Will Perplexity publish accuracy or speed results showing that agent-to-agent checking beats a single lead model?
- How many credits will a Multiagent task use out of the 10,000 included in Max?
- Will Multiagent reach plans below Max, or stay web-only?
Try it
Next time a chatbot gives you an answer you plan to rely on, paste it into a different chatbot and ask it to list errors, shaky assumptions and unsourced claims. Then verify the disagreements yourself.
Sources and links
- ReportingMarkTechPost: the article we sourced
First reported by MarkTechPost. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.