The session covered how AI agents function, first by inquiring about context—querying databases, reading files, or searching the Internet—and then performing actions based on the gathered information. The workflow typically involves a user prompt, which is sent to a Large Language Model (LLM). The LLM determines if specific tools are needed to fulfill the request. If tools are required, the agent utilizes them; otherwise, it provides a final response directly. This spectrum of actions ranges from simple queries to highly autonomous operations, highlighting the versatility of AI agents in various applications.
Integration with platforms like ChatGPT or custom GPTs necessitates defining the specific tools these agents can access. A critical consideration for these integrations is that all data must be publicly accessible, as OpenAI retains both input and output information, raising security concerns. InterSystems IRIS, however, can function as a secure tool within this framework, providing a production configuration complete with operations metrics, namespaces, and user management. This enables the creation of an IRIS agent that can be accessed via ChatGPT while maintaining data security and control.
The session also touched upon Anthropic Claude and the Model Context Protocol (MCP), which enables the use of MCP tools from a local server. This approach allows for securely grabbing data from production environments and returning it to a local instance of Claude, facilitating more complex and secure agentic workflows entirely on-premises, with LLM interactions handled via API calls.
A compelling demonstration followed, showcasing a multi-agent workflow example, which illustrated a project manager leading a team of specialized analyst tools. This team included a contract analyst, a risk analyst, a recommendation consultant, and an action planner (each of these roles is filled by an AI agent). The agents leveraged embedded Python within IRIS. For instance, when a contract is fed into the system, multiple agents collaborate to analyze it and generate a list of tasks to be performed. This collaborative approach utilizes IRIS Vector Tools, such as the IRIS Vector Search Tool, which agents call upon for efficient data retrieval. Message schemas feed prompts to the agents, and the system can even leverage external tools like Trello. Each agent is assigned a specific role, goal, and a backstory to clearly define its purpose and function within the workflow.
The multi-agent workflow further details how the first agent reads the contract and categorizes its clauses into standard versus unique elements. The second agent then focuses on the non-standard clauses, performing Internet searches to identify potential risks, citing sources, and providing justifications for each clause. Upon completion, a recommendation agent synthesizes all this information to generate actionable recommendations.
The sample workflow includes human oversight, where the system requests specific feedback on problematic contract clauses. If a human user approves, the recommendations are then automatically transferred to Trello to create actionable cards for team members. These Trello cards are comprehensive, including background information, a description of the nature of the issue, potential impact, rationale for the impact, sources, and recommended language changes to mitigate identified risks. This entire integration leverages InterSystems IRIS’ robust capabilities in data management, interoperability, and analytics, providing a comprehensive and secure solution for complex document analysis and task management.
