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

Ivan Leo introduces Manus AI, describing it as an action engine capable of executing tasks, automating workflows, and extending human reach. He highlights Manus 1.5's improvements in speed, quality, and architecture, noting the challenges of building AI agents at scale. The workshop will focus on using the new Manus API to reproduce their original Slackbot.

Manus aims to be a general AI agent meeting users where they are, through its web application, Slack app, API, browser operator, iOS app, and even a new Microsoft 365 integration for tasks like editing presentations or fixing spreadsheets. Four demos will showcase Manus's capabilities.

Ivan demonstrates a French language practice application built with Manus. Users can input daily thoughts, and Manus uses a language model to identify and correct grammar mistakes, provide translated definitions, and even generate audio pronunciations. The app also builds a user profile based on interactions.

The second demo showcases the new browser operator. Ivan prompts Manus to find a simple coffee shop near his location in New York, and Manus kickstarts a browser tab, uses Google Maps, and identifies suitable options. This remote browser operation is useful for interacting with authenticated platforms like LinkedIn or Instagram.

Ivan demonstrates an application he built to organize AIE conference events. Manus scrapes event details from the website, stores them in a JSON file, and displays them on a custom website. Users can star events, add them to their Google Calendar, and view similar events, all built by prompting Manus.

The workshop shifts focus to the Manus API. Ivan explains that the API offers the same capabilities as the web app, including sandbox execution for custom code. He details how to obtain and use a Manus API key, discussing the two models available: Manus 1.5 for complex tasks and 1.5 Light for faster, simpler queries. He demonstrates creating a basic task and polling its status.

Ivan explains three ways to provide context to Manus: uploading files, using publicly available URLs, and encoding images in Base64. He demonstrates uploading a JSON file of Rick and Morty characters to build a visualization website for internal tools, and using a Warren Buffett investor letter PDF to analyze and summarize content.

Webhooks are introduced as a more scalable alternative to polling for managing multiple tasks. When a task is completed, Manus sends an API request to a specified endpoint, notifying the user. Ivan demonstrates setting up a webhook using Modal to receive notifications when Manus tasks are created or completed, which is crucial for complex applications.

The final section demonstrates integrating Manus with Slack. Ivan outlines the process of obtaining Slack API secrets and setting up a basic Slackbot using Modal. He shows how to receive and process messages from Slack, send responses, and handle file uploads within Slack, aiming to create a fully functional Slackbot.

To enable multi-turn conversations, a simple dictionary is used to store thread IDs and associated task information, similar to a KV store. This allows the Slackbot to push responses to the same task and thread. Ivan also improves the Slack UI by using Slack Block Kit to add buttons and present information more attractively.

Ivan attempts a live demo of a more complex application: processing an invoice by uploading an image of bagels. Manus uses OCR to extract details from the receipt, references company policies stored in Notion via a connector, and then provides a response on how the expenses relate to the policy. This showcases Manus's ability to integrate with external tools and perform deep research.

Ivan answers questions about educating users on Manus API use cases, suggesting starting with the web app for exploration. He addresses user privacy, affirming that Manus cannot read user chats and data is housed in the US. He also shares interesting use cases, like automating pickleball slot bookings, and discusses upcoming features like improved browser permissions and enhanced file export options.