Building Intelligent Research Agents with Manus - Ivan Leo, Manus AI (now Meta Superintelligence)

Key Concepts

• Manis as a General AI Agent: Manis is presented as a powerful, versatile “action engine” capable of automating tasks, executing workflows, and extending human capabilities beyond simple question answering.
• Manis 1.5 Improvements: The latest iteration (1.5) delivers increased speed, improved quality, and a re-architected model designed for scalability.
• Extensive Integration Ecosystem: Manis offers access through multiple channels – web application, Slack, API, browser operator, Microsoft 365, and Mail Manis – meeting users where they are.
• API-First Approach: The newly launched Manis API empowers developers to build custom integrations and applications leveraging Manis’ core capabilities.
• Simplified AI Development: Manis aims to abstract away the complexities of AI infrastructure and model management, allowing developers to focus on application logic.

Manis: A Comprehensive Overview

Introduction to Manis & Manis 1.5

Manis is positioned as a general AI agent platform, going beyond simple answers to execute tasks and automate workflows. The platform is designed to scale to millions of daily conversations, necessitating robust infrastructure, sandboxes, and a focus on reliability. Manis 1.5 represents a significant upgrade, delivering increased speed, improved quality, and a re-architected model. The core argument is that building a general AI agent is more powerful than creating verticalized, single-purpose applications. The platform aims to “meet users where they’re at,” offering access through web applications, Slack, APIs, and mobile apps.

Core Capabilities & Integrations

Manis boasts a growing ecosystem of integrations. Demonstrations showcase its ability to perform language correction, translation, web scraping, browser automation, and task execution. Key integrations include:

• Web Application: A user-friendly interface for interacting with Manis.
• Slack App: Enables task management and automation directly within Slack.
• API: Allows developers to build custom integrations and applications.
• Browser Operator: Facilitates interaction with authenticated web platforms (e.g., LinkedIn, Instagram) through a remote browser instance.
• Microsoft 365 Integration: Launched two days prior to the workshop, enabling automated editing of Word documents, PowerPoints, and Excel templates.
• Mail Manis: Automates tasks directly from a mobile inbox, such as replying to emails.

API Workflow & Technical Details

The Manis API operates through a standard request-response cycle: sending a request, receiving a task ID and URL, polling for completion, and retrieving the results. Web hooks can be registered to receive notifications upon task completion, enabling asynchronous processing and reducing the need for constant polling. File uploads (JSON, PDFs) are supported, with files automatically deleted after 48 hours for data privacy. Key technical terms include API (Application Programming Interface), Web Hook, Sandboxes, Context Management, KV Caching, Docker Image, Modal, Slack Blocks, JSON (JavaScript Object Notation), MIME Type, and Embedding. API usage is billed at the same rate as equivalent chat interactions within the Manis web application.

Practical Implementation: Slack Integration

The workshop extensively demonstrates building a Slack bot using the Manis API. This involves receiving messages from Slack via webhooks, parsing them as tasks, and sending responses back to the appropriate Slack channel and thread. Utilizing Slack’s Block Kit enhances the user experience with richer UI elements like buttons and formatted messages.

A crucial aspect of the Slack integration is enabling multi-turn conversations and managing conversation state. A simple dictionary is used to track whether a thread has been seen before, routing subsequent messages to the correct task. More robust solutions like Moto’s dictionary, KV stores, or databases are suggested for production environments.

Real-World Applications & Use Cases

Several real-world applications were demonstrated:

• French Language Practice App: A personalized learning tool that corrects grammar, provides definitions, and adapts to user preferences.
• AI Event Scraper: Scrapes data from event websites and integrates it into a personalized timeline.
• Invoice Processing: Analyzes invoices, extracts details (merchant, amount, dates), and compares them against company policy.
• Pickle Ball Slot Booking: Automates the process of securing pickle ball court reservations by scraping a government website.
• Conference Website Creation: Demonstrates Manis’s ability to generate code and handle complex web development tasks, resulting in a website for a conference with “70 plus talks.”

Conclusion

Manis presents a compelling vision for the future of AI agents – a general-purpose platform capable of automating a wide range of tasks and integrating seamlessly into existing workflows. The newly launched API empowers developers to build custom applications, while the platform’s focus on scalability and user-centric design positions it for continued growth. The workshop highlights the potential of Manis to simplify AI development and unlock new possibilities for automation and productivity. While acknowledging current limitations (e.g., lack of persistent memory, browser automation challenges), the demonstrated capabilities and ongoing development suggest a promising future for this versatile AI agent platform.