Manus launches Wide Research to scale agent workflows

If you have ever asked an AI agent to analyze every company in the Fortune 500 or compare dozens of graduate programs, you know how painful the experience usually is. A single agent clicking through tabs one by one either runs out of context, starts hallucinating halfway through, or simply times out.

Manus wants to fix that by throwing brute-force parallel compute at the problem. The team announced Wide Research, a new system feature designed to tackle massive research tasks that require collecting and analyzing information across hundreds of entities simultaneously.

Swapping roleplay for real parallel compute

Manus has always operated differently from standard chatbots: every session runs inside a dedicated, Turing-complete cloud virtual machine. Over the past few months, the company optimized its large-scale virtualization setup to figure out how to scale compute available to a single user by 100 times.

Wide Research is the first result of that effort. Instead of relying on a single worker or traditional multi-agent setups that assign rigid roles like manager, researcher, and writer, Wide Research uses a different approach. Every subagent spun up during a task is a fully capable, general-purpose Manus instance.

When you ask for an extensive market breakdown or a massive product audit, the system spins up multiple independent VM instances in parallel. Each agent can browse, execute code in its sandbox, and pull data independently before synthesizing the final output.

Wide Research is rolling out first to Pro users, with access expanding to Plus and Basic tiers down the line.

Why it matters

We think this is one of the smartest product shifts in the autonomous agent space this year. Instead of waiting for frontier labs to magically solve long-context retention or reasoning drift, Manus is treating agent scaling as a distributed systems problem.

For knowledge workers, equity researchers, and founders, this bridges the gap between a quick summary and actual legwork. Having a cluster of sandbox-backed agents crawl hundreds of sources in parallel makes high-volume research genuinely viable through natural language. If you routinely spend days compiling competitive landscapes, testing this on a Pro plan is an easy call.

However, spinning up dozens of full virtual machines and running concurrent agent loops comes with massive infrastructure and token overhead. Gating it behind the Pro tier shows how expensive this compute is, and Manus will have to prove that these parallel runs can maintain high accuracy without burning a hole in their unit economics.

It also sends a clear signal to rivals like OpenAI and Anthropic: the real differentiator for agents might not be raw model intelligence alone, but the orchestration layer that coordinates an entire supercomputer cluster behind a simple prompt.