Simplifying AI Agent Creation and Tooling for R&D Teams
This article discusses Forter's approach to enabling 200 R&D team members, including non-engineers, to build AI agents within two weeks. Key to their success was developing an in-house Multi-Agent Communication Protocol (MCP) server for tool management, sidestepping complex RAG setups by leveraging existing enterprise search, and providing both no-code and code-based platforms for agent invocation. The strategy focused on ease of use, rapid tool integration, and addressing organizational roadblocks like security and legal early on.
Read original on InfoQ CloudForter successfully enabled a large, diverse R&D team to become "agent creators" by simplifying the underlying technical complexities. Their approach highlights how focusing on developer experience and abstracting away the 'hard parts' of AI agent development can accelerate adoption and innovation within an organization.
The Role of a Centralized MCP Server for Tooling
A core architectural decision was to build an in-house Multi-Agent Communication Protocol (MCP) server, dubbed 'Toolchain'. This server acted as a centralized registry and API gateway for all tools an AI agent could utilize. It provided a unified UI for discovering tools, managing agent connections, and configuring tool access. This abstraction was crucial for simplifying tool integration and governance.
- Ease of Tool Creation: Developers could easily add new tools by cloning a single repository and defining a simple configuration file with a description and input schema. This led to a rapid increase from 20 to nearly 100 tools in a few weeks.
- Access Control and Governance: The MCP server allowed for precise control over which tools an agent could access, preventing confusion and potential misuse. It also provided telemetry for token consumption and tool usage.
- Agent-Friendly Descriptions: Tools included clear descriptions for agents on how to use them, along with input schemas and examples, which are vital for agents to effectively interact with various services.
Sidestepping Complex RAG Implementations
Instead of building a complex Retrieval Augmented Generation (RAG) system from scratch, Forter leveraged an existing enterprise search tool, Glean. This tool was already proficient at indexing and searching across various internal data sources (Confluence, Jira, Slack, Salesforce). They exposed Glean's capabilities through three distinct tools in Toolchain:
- Search Tool: Allows agents to query internal documents and retrieve relevant snippets, adapting a human-centric search for agent use.
- Read Tool: Enables agents to retrieve the full content of a useful document.
- Summarize Tool: Provides a parameterizable summary of documents, allowing agents to focus on specific questions or topics. This significantly reduced the complexity and development time associated with providing context to agents.
Platform for Agent Invocation
Forter recognized the need for diverse agent invocation methods (interactive chat, event-triggered, scheduled). For interactive chat agents, they offered two primary experiences: a no-code solution using LibreChat and a code-based option. LibreChat was adapted to connect to multiple instances of their Toolchain, each exposing a subset of tools, allowing users to compose their agent's capabilities easily. The transparency in showing tool usage and parameters was critical for user understanding and debugging.