Just do it. (let your tools think for themselves) - Robert Chandler

Key Concepts

• Agentic MCPs (Machine-Centered Processes): Tools with increased agency and reasoning capabilities.
• Toolboxes: Collections of agentic MCPs that can be enabled or disabled for specific tasks.
• Context Pollution: The degradation of LLM performance due to excessive or irrelevant information in the context window.
• High-Level APIs: Simple, natural language interfaces for interacting with tools.
• Reliability, Reusability, and Alignment: Key characteristics of effective agentic tools.

The Problem with Traditional MCPs

• Low-Level Wrappers: Traditional MCPs are often simple wrappers around APIs not designed for LLMs, leading to messy JSON responses and context pollution.
  • Example: Sending a Slack message using an MCP resulted in listing all users and channels, confusion, and ultimately a message in the general channel.
• Tiny Scope: MCP tools are often wrappers around individual functions, making it difficult for LLMs to reason over multiple calls and compose complex tasks.
• Multi-Page Pagination: APIs that require looping over results pollute the context window and increase the reasoning burden on the LLM.
• Authentication: Managing API keys and bot creation can be cumbersome.
• Tool Overload: Adding many tools increases noise in the context window, making it harder for the agent to perform effectively.
  • Example: Adding Slack and Notion MCPs can introduce a large number of tools, which can be overwhelming.

The Solution: Agentic MCPs

• Increased Agency: Tools should be more like specialized agents with the ability to reason and perform complex tasks independently.
  • Analogy: Instead of a T-Rex holding a tiny spanner, think of a team of Avengers with specialized roles (Hulk for smashing, Hawkeye for precision).
• Blurring the Line: The distinction between tools and agents should be blurred, with agents using other agents as tools.
• Tidy, Simple APIs: Agentic MCPs should have natural language APIs that provide reliable, reusable, and high-quality outputs.

Wordware's MCP Toolbox

• Turning Wordwares into Tools: The toolbox allows users to create agentic MCPs from their Wordwares.
• Example: Competitor Analysis: A Wordware is used to perform competitor analysis by scraping Twitter and writing a summary to Notion. This requires taste, reasoning, and integration with multiple platforms.
  • The Wordware is customized with specific details about the user's company and desired analysis.
  • The output is a nicely formatted summary in a Notion page.
• Toolbox Functionality:
  • Multiple tools can be grouped into a toolbox.
  • Toolboxes can be enabled or disabled for different tasks.
• Integration with LLMs: The toolbox allows agents like Claude to use Wordware tools.
  • The user can grant permission for the agent to use the tool.
  • The agent can then perform the task using the Wordware.

Demonstration

• Creating a Competitor Analysis: The presenter demonstrates creating a competitor analysis for Anthropic AI using the Wordware toolbox and Claude.
• Output: The analysis is written to a Notion page with a summary of Anthropic's tweets, highlighting their focus on tweeting style.

Key Takeaways

• Agentic MCPs can significantly improve the reliability, reusability, and alignment of AI agents.
• By giving tools more agency and providing them with high-level APIs, we can reduce context pollution and improve the reasoning capabilities of LLMs.
• Wordware's MCP toolbox provides a platform for building and managing agentic MCPs.
• The analogy of the Avengers highlights the importance of specialized tools and agents working together to achieve complex goals.