The Next Breakthrough In AI Agents Is Here

Key Concepts

Multi-agent AI system, task decomposition, chain of thought injection, AI agents, wrappers, foundational models, tool integrations, proprietary evaluations, sustainable differentiation.

Manis: A New Agentic AI Platform

Overview

Manis is a new AI platform developed by a Chinese startup, designed as a general-purpose AI agent. It's a multi-agent AI system that aims to complete various tasks, from travel planning to financial analysis, by coordinating a team of sub-agents.

How Manis Works

1. Input: The user provides a prompt.
2. Planning: A "planner agent" breaks down the task into manageable subtasks, creating a master plan.
3. Execution: Sub-agents, each with a specific domain (knowledge, memory, execution), are assigned subtasks. Manis has 29 integrated tools for tasks like web navigation, code execution, and information retrieval.
4. Tool Selection: Sub-agents intelligently choose the appropriate tools for their tasks.
5. Synthesis: An "executor agent" combines the outputs from the sub-agents into a final synthesized output for the user.

Technical Details

• Dynamic Task Decomposition: Manis uses a sophisticated algorithm to break down complex instructions into clear execution paths.
• Chain of Thought Injection: This technique allows agents to actively reflect and update plans, ensuring stability.
• Underlying Model: Manis uses Anthropic's Claude 3.7 Sonnet.
• Cross-Platform Execution: Manis integrates with open-source tools like YC company Browser use and E2B's secure cloud sandbox environment.

Capabilities and Applications

Manis can handle a wide range of real-world tasks, including:

• Creating travel itineraries
• Performing detailed financial analyses
• Generating educational content
• Compiling structured databases
• Comparing insurance policies
• Sourcing suppliers
• Assisting with presentations

Performance Benchmarks

• Gaia Benchmark: Manis achieved a score of 86.5% on the Gaia benchmark, which tests AI agents on reasoning, multimodal handling, web browsing, and tool proficiency. Humans typically score around 92%, while OpenAI's Deep Research scored 74%.

Manis as a "Wrapper" and its Implications

The "Wrapper" Argument

Some consider Manis a "wrapper" because it stitches together existing foundational models and tool calls. The video argues that many successful AI products today could be classified as wrappers. Examples include:

• Cursor and Windsurf: Integrate LLMs with external APIs and developer-focused tooling.
• Harvey: Combines foundational models with legal-specific tool integrations.

What Makes a Successful Wrapper?

Successful wrappers are characterized by:

• Intuitive UI
• Proprietary evaluations
• Careful fine-tuning of foundational models
• Thoughtfully designed multi-agent architectures

Manis' Strengths and Weaknesses as a Wrapper

Strengths:

• Lower per-task costs (around $2 per task) compared to competitors like OpenAI's Deep Research.
• Greater transparency and user control, allowing users to inspect, customize, or replace individual sub-agents and tool integrations.
• Exposing the file system so users can see exactly what the agents are doing.

Weaknesses:

• Coordination across specialized agents becomes difficult as tasks scale or complexity grows.
• Advantages (UX refinements, targeted fine-tuning, thoughtful integrations) are vulnerable to competitors.
• Vulnerable to disruption from API pricing changes or provider policy shifts.

Sustainable Differentiation for AI Products

The critical challenge is identifying sustainable differentiation. Strategies include:

• Investing in proprietary evaluations that are expensive or time-consuming to replicate.
• Embedding workflows deeply into specific user routines to increase switching costs.
• Identifying integrations with platforms or datasets competitors can't easily access.

Quote: "From day one, they decided to work orthogonally to model development, wanting to be excited rather than threatened by each new model release" - Yichchow Peak G, Manis Co-founder

Conclusion

Manis represents a significant step forward in the development of usable AI agents. While it may be considered a "wrapper," its multi-agent architecture, focus on transparency, and impressive benchmark performance demonstrate its potential. The key to long-term success lies in establishing sustainable differentiation through proprietary evaluations, deep user integration, and exclusive partnerships. The video concludes that success in AI depends on effectively integrating existing models into products that users find valuable.