TDM Agent Architecture

The TDM Agent runs entirely inside the customer's K2view Fabric deployment, beside TDM, on the AI Fusion agent runtime. It only prepares TDM task executions — a person always reviews and runs them. This article describes how the agent is built on top of Fabric and AI Fusion, and how a request flows through it.

For the data sent to the LLM and the security controls, see TDM Agent Security and Privacy.

Core Control Principle: It Prepares, It Never Executes

The agent will The agent will not
Reuse existing, administrator-defined TDM tasks Create new tasks from scratch
Prepare a task execution with overrides Start an execution — a user clicks the link
Ask when the entity, environment or action is ambiguous Guess an environment from its type
Explain each condition it will select by Show raw override JSON or internal IDs

Every data movement therefore still passes through TDM's existing task permissions, environment permissions and execution controls.

This principle applies to the TDM Agent in chat and through the REST API. The TDM MCP tools are a separate channel: an MCP client can also start a task execution with the StartTaskExecution tool, under the same TDM permissions and execution validations.

Components

The TDM Agent is a set of AI Fusion agents and Broadway tool flows deployed on the customer's own K2view Fabric, beside TDM. It adds no new external service. The only outbound call is to the LLM provider the customer configures.

TDM Agent architecture on K2view Fabric

Requests enter only through authenticated Fabric web services and MCP. The agents reach TDM only through allow-listed tool flows. Data moves between environments only through a TDM execution: started by a person from the prepared draft, or by an MCP client through the StartTaskExecution tool.

Layer Components
Clients TDM chat UI; CI or test-management tools through the REST API; MCP clients
Access Fabric web services and MCP servers, Fabric authentication and RBAC
Agents Specialized AI Fusion agents for request handling, task search and override building, deployed in the TDM Data Product
Tools A fixed set of tools to search tasks, read task and parameter definitions, check execution status and prepare a draft execution (allow-listed per agent)
TDM Task catalog, environments, draft executions, permissions, execution engine
Platform K2view Fabric: flow engine, background jobs, secret manager, LLM interfaces
Storage Conversation traces, session records, API requests and draft executions, all in the customer's Fabric and TDM databases

Agentic Flow

Each chat or REST API request passes through a fixed chain of specialized agents. A trusted hook sets the scope, the LLM works only through allow-listed tools, and the run always ends at a draft that a person must start.

TDM Agent agentic flow

In chat, when several good matches exist, they are shown to the user to choose from. Through the REST API, each test case runs as its own audited session and never asks questions. It ends READY with a review link, or FAILED with the agent's reasoning.

Previous

TDM Agent Architecture

The TDM Agent runs entirely inside the customer's K2view Fabric deployment, beside TDM, on the AI Fusion agent runtime. It only prepares TDM task executions — a person always reviews and runs them. This article describes how the agent is built on top of Fabric and AI Fusion, and how a request flows through it.

For the data sent to the LLM and the security controls, see TDM Agent Security and Privacy.

Core Control Principle: It Prepares, It Never Executes

The agent will The agent will not
Reuse existing, administrator-defined TDM tasks Create new tasks from scratch
Prepare a task execution with overrides Start an execution — a user clicks the link
Ask when the entity, environment or action is ambiguous Guess an environment from its type
Explain each condition it will select by Show raw override JSON or internal IDs

Every data movement therefore still passes through TDM's existing task permissions, environment permissions and execution controls.

This principle applies to the TDM Agent in chat and through the REST API. The TDM MCP tools are a separate channel: an MCP client can also start a task execution with the StartTaskExecution tool, under the same TDM permissions and execution validations.

Components

The TDM Agent is a set of AI Fusion agents and Broadway tool flows deployed on the customer's own K2view Fabric, beside TDM. It adds no new external service. The only outbound call is to the LLM provider the customer configures.

TDM Agent architecture on K2view Fabric

Requests enter only through authenticated Fabric web services and MCP. The agents reach TDM only through allow-listed tool flows. Data moves between environments only through a TDM execution: started by a person from the prepared draft, or by an MCP client through the StartTaskExecution tool.

Layer Components
Clients TDM chat UI; CI or test-management tools through the REST API; MCP clients
Access Fabric web services and MCP servers, Fabric authentication and RBAC
Agents Specialized AI Fusion agents for request handling, task search and override building, deployed in the TDM Data Product
Tools A fixed set of tools to search tasks, read task and parameter definitions, check execution status and prepare a draft execution (allow-listed per agent)
TDM Task catalog, environments, draft executions, permissions, execution engine
Platform K2view Fabric: flow engine, background jobs, secret manager, LLM interfaces
Storage Conversation traces, session records, API requests and draft executions, all in the customer's Fabric and TDM databases

Agentic Flow

Each chat or REST API request passes through a fixed chain of specialized agents. A trusted hook sets the scope, the LLM works only through allow-listed tools, and the run always ends at a draft that a person must start.

TDM Agent agentic flow

In chat, when several good matches exist, they are shown to the user to choose from. Through the REST API, each test case runs as its own audited session and never asks questions. It ends READY with a review link, or FAILED with the agent's reasoning.

Previous