AI agents are everywhere, powering chatbots, automating tasks, and handling workflows. But there is a fundamental shift happening. I recently built a drag-n-drop no-code agent builder that works off of a workflow you create using open source LLM models, Node.js, ReactFlow, Ollama, and Langchain, purely to learn how agents work and the technology to create them. At the time, I thought it was very cool to drag nodes and connect through prompt chaining to see agents stepping through workflow tasks and passing data between them.
Enter Agentic AI, a new paradigm where AI does not just execute tasks. It understands, adapts, and orchestrates dynamically. After watching several testing videos of the new Manus, a general Agentic AI system built in China, I was blown away. Manus is built built on open source tools using Claude 3.5 Sonnet and two fine-tuned Qwen models and in private invite-only beta. In just two weeks after working on my “very cool” drag-n-drop agent builder using defined workflows, I realize now it is already obsolete.
AI agents are like specialized workers. They perform predefined tasks based on rules or prompts. While useful, they are limited by static behavior.
Agentic AI is different. It’s goal-driven. Instead of just responding to prompts, it autonomously plans, reasons, and adapts. It breaks down complex objectives into smaller steps, coordinates multiple agents, and adjusts strategies in real time.
Manus is not just another AI agent. It is an orchestrator of agents. It dynamically generates and coordinates agents, ensuring they work together to achieve high-level goals.
Imagine giving Manus a broad directive like improving lead generation for a sales team. Instead of relying on static workflows, it would:
- Break the goal into sub-tasks like prospecting, outreach, follow-ups, and tracking
- Generate specialized agents for each task such as a research agent for prospecting, a messaging agent for personalized emails, and an analytics agent to track responses.
- Adapt strategies based on real-time feedback, adjusting messaging, timing, and targeting
- Continuously refine its approach by analyzing successful conversions, outreach sequences, and better targeting
Instead of following fixed rules, Manus dynamically orchestrates agents to keep optimizing the process.
**UPDATE ONE DAY AFTER WRITING THIS POST** Just when you thought AI wasn’t moving fast enough, a team of four developers from MetaGPT in China recreated Manus as OpenManus available as open source on Github. However, after first look, it doesn’t appear to be as comprehensive in its results as Manus. Identical tests performed on Manus with OpenManus are not as complete, but it’s an interesting start for a new framework.
