First Hands‑On Exploration of OpenManus: Installation, Architecture, and Real‑World Tests

Installation

OpenManus can be installed via several methods; the article uses Conda to create a virtual environment and then clones the repository:

conda create -n open_manus python=3.12
conda activate open_manus
git clone https://github.com/FoundationAgents/OpenManus.git
cd OpenManus
pip install -r requirements.txt

Optionally, the Playwright browser automation tool can be installed with playwright install. The configuration file config.toml is edited to select the DeepSeek LLM and the Qwen2.5‑VL‑72B‑Instruct visual model.

Overall Architecture

OpenManus follows the ReAct (Think‑Act‑Observe) paradigm and is built with an inheritance hierarchy: BaseAgent, ReActAgent, and ToolCallAgent. Users can extend BaseAgent to create custom agents, and multi‑agent support is available via configuration.

The three ReAct stages are:

Think : generate reasoning based on system prompts, memory, and user input; decide which tools to call and with what parameters.

Act : execute the selected tool calls, process results, and store them in the agent’s memory.

Observe : analyse tool output, update the agent’s state, and decide whether to continue or terminate.

Case Tests

Test 1 – GitHub statistics : The user asks OpenManus to retrieve the star and fork counts of the OpenManus repository and plot them with Python. The agent proceeds through 20 steps, starting by opening a browser with the browser_use tool, navigating to the project URL, locating the star/fork labels, and attempting to scrape the numbers. The article notes that the scraped values are inaccurate due to limitations of browser_use. Finally, OpenManus generates a Python script, runs it, and produces the plotted result.

Test 2 – Physics animation : The user requests a series of clear animation slides for a middle‑school lesson on momentum conservation, packaged as an HTML presentation. OpenManus completes the task in four steps, directly generating the HTML file. Unlike other agents, it then verifies each animation segment, using browser_use to test the HTML and ensure correctness. The resulting animated presentation is displayed and judged to be satisfactory.

Summary

The author concludes that OpenManus performs well overall, especially thanks to its ReAct‑driven stepwise execution, which cleanly decomposes complex tasks. The animation test showcases the agent’s ability to produce and validate rich output. However, the first test reveals that the browser_use tool can mis‑capture web elements, leading to incorrect data, indicating an area for future improvement.