On 31 July 2025, the AI startup Manus released Wide Research—a new feature that deploys more than a hundred autonomous agents on a single task. Instead of relying on a single model to sequentially crawl, extract and synthesize information, Wide Research orchestrates a swarm of general‑purpose Manus agents that work in parallel. This shift from deep, serial reasoning to wide, parallel collaboration promises to turn large‑scale information gathering into a task measured in minutes rather than days.
From Deep to Wide: Solving the Sequential Bottleneck
Traditional agent tools such as OpenAI’s Deep Research and similar offerings from Google or xAI build deep knowledge by having one high‑capacity agent step through sub‑tasks sequentially. This approach excels when the problem space is narrow and demands nuanced reasoning. However, when a researcher needs to compare hundreds of products or gather data on dozens of organizations, a single agent becomes a bottleneck; tasks must be completed one after another, and the agent may lose context across a very broad scope.
Manus addresses this limitation by spawning up to 100 fully autonomous agents per user requestventurebeat.com. In a demonstration posted on X, co‑founder Yichao “Peak” Ji instructed Wide Research to compare 100 sneakers; the system instantly created 100 sub‑agents, each responsible for analysing the design, pricing and availability of a single shoe, and returned a sortable matrix in both spreadsheet and webpage formats within minutesventurebeat.com. A similar demo showed 50 agents generating poster designs simultaneouslywinbuzzer.com. These examples illustrate how parallelism eliminates the sequential bottleneck inherent in deep research systems.
Architectural Innovation: Parallelism and Collaboration
Wide Research relies on a large‑scale virtualization infrastructure. Every Manus session runs inside its own virtual machine (VM), giving users a personal cloud computing environment that is Turing‑completemanus.im. When Wide Research is invoked, this infrastructure scales compute power “by 100 ×,” effectively turning a single VM into a cluster of coordinated machineswinbuzzer.com.
Manus emphasises that the feature is more than just having lots of agents—it is about agent‑to‑agent collaboration. Each sub‑agent is a fully featured Manus instance, not a specialized manager or coderwinbuzzer.com. Agents communicate through a proprietary protocol that allows them to divide tasks, share partial results and converge on a coherent output. This system‑level mechanism for parallel processing and agent collaboration is what Manus claims differentiates Wide Research from other multi‑agent systemsindianexpress.com. Because every agent is general‑purpose, the system can pivot across domains without pre‑defined roles or rigid templates, letting users ask for almost anything—from ranking MBA programmes to analysing open‑source toolsmanus.im.
Pricing and Availability
Manus initially restricted Wide Research to users on its Pro plan, which costs US $199 per month and includes 19 900 monthly credits. The company plans to roll the feature out gradually to the Plus (US $39 per month) and Basic (US $19 per month) tiers as wellwinbuzzer.com. Free users are limited to a single concurrent agent and cannot access the new featureventurebeat.com. The high price reflects the computational cost of orchestrating hundreds of high‑performance AI agents across cloud providers. Manus uses Anthropic’s Claude models to power these agentsventurebeat.com and runs workloads on both Google Cloud and AWS infrastructures.
Wide vs. Deep: Key Differences
The distinction between Wide and Deep research can be summarised along several dimensions:
| Dimension | Wide Research | Deep Research |
|---|---|---|
| Processing architecture | Parallel – spawns dozens or hundreds of general‑purpose agents that work simultaneously on sub‑taskswinbuzzer.com. | Sequential – a single agent steps through sub‑tasks in order. |
| Task flexibility | Agents are general‑purpose; tasks aren’t tied to rigid roles or domainsindianexpress.com. | Single agent follows a fixed workflow; less adaptable when the task spans diverse domains. |
| Collaboration mechanism | Proprietary agent‑to‑agent protocol distributes work and merges resultswinbuzzer.com. | No multi‑agent collaboration; the lone agent composes the final report. |
| Scalability | Scales compute to 100 × via VM clusters, enabling high‑volume research in minuteswinbuzzer.com. | Limited by the speed and capacity of a single model; high‑volume tasks take much longer. |
| Ideal use cases | Comparing many products or companies, exploring large datasets, generating multiple creative assets simultaneouslyventurebeat.com. | Deep dives into complex topics where nuance and stepwise reasoning are paramount. |
Real‑World Applications
Wide Research’s ability to spin up large agent swarms opens up new workflows:
- High‑volume market or product research. Users can ask the system to compare hundreds of products, such as the top 100 sneakers or 1 000 stocks; each item is assigned to a different agent and results are aggregated into a single matrixventurebeat.com. This dramatically reduces the time required compared with sequential methods.
- Creative design exploration. Wide Research can generate dozens of design variations at once. In one demo, 50 agents created poster designs in distinct styles, packaged as a downloadable ZIP filewinbuzzer.com.
- Corporate and academic surveys. The tool can compile information on all Fortune 500 companies, rank MBA programmes or synthesise insights from hundreds of research papers, tasks that previously demanded lengthy manual workindianexpress.com.
- API and data integration. Because agents operate in a Turing‑complete VM, they can interact with APIs, run code and fetch data autonomouslymanus.im, making Wide Research useful for developers who need to integrate diverse data sources into a single analysis.
Challenges and Criticisms
Although Wide Research showcases a bold architectural vision, several issues temper the enthusiasm:
- High cost and resource use. The Pro tier’s US $199 monthly fee is steep, and spinning up hundreds of agents consumes significant compute. Critics worry that such costs put the tool out of reach for individual researchers or small teamsmedium.com.
- Lack of benchmarks. Independent reviewers note that Manus has not released performance comparisons or technical details showing that a swarm of sub‑agents outperforms a single high‑capacity agentventurebeat.com. Without transparent metrics, the practical benefits remain unprovenwinbuzzer.com.
- Coordination complexity. Running many agents increases the risk of inconsistent or redundant outputs; Manus has not disclosed how its protocol prevents conflicting resultsmedium.com. Early user feedback also highlights occasional crashes and server overloadsmedium.com.
- Regulatory scrutiny. Wide Research’s fully autonomous operation has attracted regulatory attention. U.S. states such as Tennessee and Alabama banned Manus across state networks due to concerns about censorship, propaganda and security vulnerabilitieswinbuzzer.com. These bans underscore the broader debate over how much autonomy AI agents should have.
Outlook
Wide Research represents a meaningful step toward scalable AI workflows that combine natural‑language interaction with super‑computing levels of parallelism. By transforming a single VM into a cluster of cooperating agents, Manus seeks to democratise high‑volume research and creative exploration. Early demonstrations show impressive speed and breadth, and the underlying infrastructure hints at future capabilities beyond research, such as software testing or large‑scale data synthesis.
Yet the technology is still experimental. Its computational demands raise cost barriers, and the absence of transparent benchmarks makes it difficult to judge real‑world benefits. Whether Wide Research becomes a staple of daily work depends on Manus’ ability to refine its collaboration protocols, provide clearer performance data and address regulatory concerns. Nevertheless, the launch of Wide Research marks an important moment in the evolution of AI agents—showing that the next wave of productivity tools may be characterised not by deeper thinking alone, but by wider collaboration at scale.
