Whitepaper
Overview of Agentic AI and the Insight Workplace

Overview

The Mindbreeze Insight Workplace represents a paradigm shift in how enterprises interact with their corporate knowledge, moving beyond passive search retrieval toward an active, Agentic AI ecosystem. At the core of this system is the transition from a single-prompt interface to a sophisticated multi-agent architecture, designed to orchestrate complex workflows and deliver high-precision, actionable insights.

The architecture is governed by the Mindbreeze Insight Workplace Root Agent, which acts as the central intelligence and orchestrator. Rather than attempting to solve every request with a general-purpose model, the Root Agent employs a sophisticated reasoning loop: it analyzes the user input, selects the most appropriate specialized sub-agent, and manages a recursive handoff process. This ensures that every query is routed to the domain expert best equipped to handle it, with a final evaluator stage that validates the output for accuracy and relevance before it reaches the user.

This ecosystem leverages a diverse array of specialized Sub-Agents to handle distinct business functions, like the Expert Finder Agent which connects users with the right human expertise within the organization.

The Insight Touchpoint Generator Agent synthesizes complex information into usable touchpoints for business processes.

Additionally, the Mindbreeze Insight Workplace also offers A2A (Agent-2-Agent) Integration which extends the Insight Workplace's capabilities by seamlessly interfacing with third-party agents.

The power of these agents is derived from their integration with a robust set of tools. By leveraging the Mindbreeze InSpire AI Search and the underlying AI Index Schema & Ontology, the agents operate with a deep semantic understanding of the organization's unique data structures. Furthermore, the integration of the Model Context Protocol (MCP) allows these agents to move beyond information retrieval and actively engage with tools, enabling a truly agentic workflow where the Insight Workplace can not only find information but execute tasks across the corporate landscape.

By unifying these components, the Mindbreeze Insight Workplace transforms the corporate search experience into a collaborative environment where Agentic AI handles the complexity of discovery, synthesis, and execution, allowing the user to focus on high-value decision-making.

Root Agent and Sub-Agents

Root Agent

The root agent (orchestrator) receives the initial user request. The root-agent can automatically delegate (auto flow) a task to a specialized sub-agent. The root agent considers not only the instruction and tools but also the description of each sub-agent. If the LLM determines that the user input can be better handled by a sub-agent’s described capability the control will automatically be transferred to the sub-agent for that turn. If needed one can also build routed agents.

The task of the root agent is given the user query and memory (e.g. the Insight Journey) provide an Insight Touchpoint that provides an interactive experience and automates many steps that the user manually would need to perform. An Insight Touchpoint can be any Insight App such as returning Trusted facts and Answers or automatically generating Documents given the Information in Mindbreeze, available tools and sub-agents.

Sub-Agents

Sub-agents are specialized agents, each designed for a specific capability. Sub-agents have a highly focused instruction and a clear description. With the description the root agent uses to decide when to delegate to these sub-agents.

Agent-2-Agent (A2A) Protocol

The A2A protocol can be used to integrate agents into the Insight Workplace. It uses JSON-RPC 2.0 over HTTP(S) and the gRPC protocol. The fundamental unit of work is the Task. Tasks are identified by unique ID, are stateful and progress over time.

Tools

Tools extend the ability of agents by, given the tools description and the input and output specification, the LLM agent, instead of returning an answer directly, is able to call the tool which then places the tool output in the conversation. This output is then additional context, and can also be the final response.

Mindbreeze Insight App Modules

Existing Insight App Modules and additional Modules built with the designer, are effective Tools. The Insight Touchpoint Generator agent can make use of these tools to generate Insight Touchpoints by using these tools at hand. Insight Touchpoints can also be parameterized which can be resolved by the LLM agent.

Mindbreeze InSpire AI Search Tools

The Mindbreeze InSpire AI API is accessible as a tool and exported using via an MCP server endpoint. The API tools describe the input and output schema.

External Model Context Protocol (MCP)

External MCP servers allow the AI application to access remote tools, services and data sources. Mindbreeze InSpire provides built-in API tools to use the AI search that provides access to enterprise data sources in a secure and efficient way. For use cases such as workflow automation, extended application context, etc. external tools can be used.

Modules in Insight Workplace

Insight App Widgets and Modules

Insight App Modules are the fundamental building blocks of Insight Touchpoints in the Insight Workplace. Insight Modules can be nested and reused across different Insight Touchpoints and Insight Apps.

An Insight Touchpoint is typically composed of multiple Insight App Modules. For example, a 360-degree view of business information may consist of dedicated Insight App Modules for persons, documents, activities, or other domain-specific entities.

For detailed overview of all available Insight App Widgets, please refer to the Available Widgets section in the manual Development of Insight apps.

Using and Building Insight App Modules for Agentic AI

Insight App Modules can be visually created and maintained using the Insight App Designer.

Insight App Modules are reusable building blocks that can be composed into larger Insight App Modules and Insight Touchpoints. In addition to defining the visual presentation of information, modules can expose metadata, parameters, and schema information that describe their capabilities and required inputs.

This metadata allows Insight Workplace Agents to discover and use Insight App Modules during agentic processing. Agents can instantiate modules with generated or retrieved data and combine multiple modules to create context-specific Insight Touchpoints for the user.

The following screenshots show the Insight App Designer with the available actions for creating, editing, and saving Insight App Modules and also how to interact with Insight App Modules.

AI Agent Telemetry and Feedback

User Feedback

Users have the option to share feedback about the quality of the results they are getting.

Administrators then get detailed information about the complete execution traces that lead to the generated response, including service level instrumentation data. This labelled data can be used as a dataset and derive measures to improve the overall quality of the generated responses.

These datasets can be automatically evaluated with automatically calculated metrics such as faithfulness and factual correctness.

For more information, please refer to the chapter “Evaluation of a pipeline” in Whitepaper - Administration of Insight Services for Retrieval Augmented Generation.

Distributed Instrumentation and Tuning

Mindbreeze InSpire services are instrumented using the multi-agent distributed telemetry product called app.telemetry. This allows to easily navigate complex execution flows using an intuitive visual telemetry UI.

As described in the Overview above, the Insight Workplace consists of multiple agents interacting with LLMs and tools. Each of these steps can be retraced in the UI, including API-Calls and resource access with user identity, to allow comprehensible tracing and debugging.

For more information, please refer to the Telemetry and Feedback section in the manual Whitepaper - Administration of Insight Services for Retrieval Augmented Generation.

Specific auto-scoring metrics for sub-agent and tool calling tasks, such as Insight App generation, are supported to optimize the agentic output given instrumentation data.

Auditing and LLM Guardrails

Auditing of Actions and Operations

Mindbreeze InSpire implements a fine granular and extensible role-based access control (RBAC) system across services. Actions being performed on services are access controlled and services are instrumented. Services act on behalf of a user are authenticated using system accounts and always need tokens of the user in order to access data.

App.telemetry log pools are the backing store of an instrumented application and are persisted in an internal database. Every log pool has a configurable retention period. Services such as syslog and auditd are used to log and audit critical events, system level and operating system (OS) level interactions such as logins or configuration changes, or web server access with the user that was used to perform the requests.

Auditing of Retrieval and Data Access

with retrieving and auditing sensitive data, provided the main ingredients to provide secure agentic AI applications from the start. Every service call that is performed by an agent always is done with the permissions of the user on which behalf the call was made. Furthermore, various audit logs and authorization scopes (Early binding, Late binding, Early binding with live reauthorization) can be performed based on the sensitivity of the data and the complexity of access checking process. The same level of security is guaranteed regardless of the retrieval method. The Query Audit Trail is kept in a specific app.telemetry log pool and guarantees that all results and answers that have been retrieved by the user are logged, regardless of the way the user has accessed them.

LLM Guardrails

The second important aspect about security in generative AI based applications is being able to control and transform the answers that are being returned from the provided context. For this the InSpire LLM API Proxy supports guardrail configuration and integration in various platforms.