Editorial illustration for Perplexity Deploys Up to 16 AI Agents for Cross-Checked Tasks
Perplexity Launches 16 AI Agents for Task Verification
Perplexity Deploys Up to 16 AI Agents for Cross-Checked Tasks
Perplexity added a new effort tier to Computer, its platform for research and multistep tasks, called Multiagent. It sits above Ultra, the previous ceiling introduced when Perplexity rolled out effort controls on September 17, 2026, alongside Light, Standard and High. Multiagent changes how the system is structured, not just how hard it works.
Instead of one orchestrator model doing the reasoning and farming out small jobs to helpers who report back individually, Multiagent puts a lead agent in charge of a team of up to 16 supporting agents running on different frontier models at once. In Perplexity's own example, Claude Fable 5.2 acts as the planner, dividing work among agents built on Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. Those agents don't just work in isolation and hand results up the chain. They talk to each other directly, comparing notes and checking one another's output before the lead agent compiles anything.
Perplexity hasn't published benchmarks, accuracy figures or speed comparisons for the new tier, and it's currently limited to the web version of Computer on the Max plan, with no local or quantized deployment option.
Perplexity Computer Multiagent is a new effort level that sits above Ultra. A lead agent runs a team of agents on different frontier models. Those agents work in parallel, share findings directly and check each other’s work.
Why this matters
Perplexity is betting that cross-checking beats single-model confidence, and that's a reasonable bet given how often one lead agent's blind spots go uncaught. Running Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon against each other, with Claude Fable 5.2 doing the planning, turns a research task into something closer to peer review. For developers building on Perplexity Computer, the real shift is structural: agents reporting sideways to each other, not just up to one lead, changes how you'd debug a failed task or trace where an error crept in.
We'd want to see the cost and latency numbers before calling this a clear win. Sixteen agents plus a lead, all on frontier models, is not cheap compute, and Perplexity hasn't said what Multiagent costs relative to Ultra. For analysts and founders evaluating agent platforms, the question isn't whether cross-checking improves accuracy.
It's whether that accuracy gain survives contact with a bill. Watch for Perplexity to publish benchmarks against single-agent Ultra runs on the same tasks.
Common Questions Answered
How does Perplexity's Multiagent effort tier differ from the previous Ultra tier?
Multiagent changes the structural approach rather than just increasing computational effort. Instead of one orchestrator model delegating tasks to individual helper agents, Multiagent deploys a lead agent managing a team of up to 16 agents on different frontier models that work in parallel, share findings directly, and cross-check each other's work.
What frontier models does Perplexity Computer Multiagent use for cross-checking tasks?
Perplexity Computer Multiagent runs Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon against each other to perform cross-checking, while Claude Fable 5.2 handles the planning and orchestration. This multi-model approach transforms research tasks into something closer to peer review.
Why does Perplexity believe cross-checking between multiple agents improves task accuracy?
Perplexity is betting that cross-checking beats single-model confidence because a single lead agent's blind spots often go uncaught. By having multiple agents on different frontier models verify each other's work, the system can catch errors and oversights that a single model might miss.
What is the key structural change in how agents communicate within the Multiagent system?
In the Multiagent system, agents report sideways to each other rather than only reporting upward to a single lead orchestrator. This peer-to-peer communication structure allows agents to share findings directly and perform mutual cross-checking instead of operating in isolation.
Further Reading
- Introducing Computer’s Multiagent effort level - Perplexity
- Perplexity Computer Adds Multiagent Mode, Where Up to 16 Agents on Rival Models Check Each Other - i10X News
- Towards a science of scaling agent systems: when and why agent systems work - Google Research
- An Empirical Study of Multi-Agent Collaboration for Automated Research - arXiv
- Perplexity Launches Multiagent to Coordinate Frontier AI Models - Unite.AI