Manus is a tool-using agent — it autonomously selects and uses the right tools for each step of a task. Web search for current information, code execution for computation, data extraction tools for structured data, and API calls for live integrations. It orchestrates tools rather than relying on a single capability.
Manus breaks the task into steps and determines which tool category is most appropriate for each: web browsing, code execution, file manipulation, or calculation.
Each tool is called when its step is reached in the execution plan. Results flow between tools — web search output becomes code input, code output becomes file content.
After each tool call, Manus checks whether the result makes sense given the expected outcome and re-tries or pivots if the result is unexpected.
The outputs from all tools are combined into a coherent final deliverable — not separate tool outputs, but one integrated result.
Combining web research with numerical analysis
Research the top 15 enterprise SaaS companies by ARR growth rate in 2023-2024. Pull the data, calculate average and median growth rates by category (infra vs app vs vertical), and visualize with a bar chart. Deliver as a report with the chart embedded.
Building a strategy document from research
Research the current state of AI regulation globally. Then write a 10-page strategic brief for our legal and compliance team covering the 5 most important regulatory developments, risk implications for a US-based AI company, and recommended compliance actions.
Full software deliverable package
Build a Python library for common statistical calculations (mean, median, std dev, correlation, linear regression). Include full type hints, comprehensive test coverage, API documentation, and a usage examples file. Deliver as a complete distributable package.
If you know the task needs multiple tools, frame it that way: "Research the data, analyze it, and deliver as a formatted report." This primes Manus to plan for tool orchestration.
Watch which tools Manus uses for your task — if it's using only web search when you expected code execution, redirect to ensure the right tools are applied.
For tasks requiring authenticated API calls (sending emails, posting to Slack, querying private databases), provide credentials explicitly in the task context.
"After completing the task, include a brief methodology section explaining which tools you used for each major step and why." This documents the research method for reproducibility.