Introduce a Structured "Workflow" Resource Type for Deterministic Execution

[rgeissen]

Is your feature request related to a problem? Please describe.
The use of free-form text prompts to define complex, multi-step, procedural tasks is a source of execution unreliability. A non-deterministic language model is ill-suited for orchestrating rigid, algorithmic processes, which can result in execution errors such as skipping steps, hallucinating new steps, or failing to terminate correctly.

Describe the solution you'd like
Introduce a new, structured resource type on the MCP server, provisionally named "workflow." A workflow would be defined in a machine-readable format (e.g., JSON or YAML) that explicitly outlines the sequence of operations. This definition would include the specific tool calls, control flow logic such as loops and conditionals, and designated points where LLM intervention for non-procedural tasks (like summarization) is required.

This allows a client application to fetch the structured workflow and use a local, deterministic engine to execute the plan, ensuring a predictable and correct outcome.

Example of a Structured Workflow Definition:

name: databaseQualityWorkflow
description: "Performs a comprehensive data quality assessment on all tables in a database."
phases:
  - name: "Get Database Tables"
    steps:
      - tool: base_tableList
        args:
          db_name: "{{db_name}}" # Input parameter
        provides: table_list
  - name: "Collect Table Information"
    loop:
      over: table_list
      as: table
    steps:
      - tool: base_tableDDL
        args:
          db_name: "{{db_name}}"
          table_name: "{{table.TableName}}"
        provides: ddl_output
      - llm_step:
          prompt: "Generate a business description for the table and columns based on this DDL: {{ddl_output.results[0].Request Text}}"
          provides: business_description
      - tool: qlty_columnSummary
        args:
          database_name: "{{db_name}}"
          table_name: "{{table.TableName}}"

Describe alternatives you've considered
The primary alternative is to continue using natural language prompts and attempt to improve reliability through advanced prompt engineering. This involves adding more constraints, rules, and few-shot examples to guide the language model. This approach is suboptimal as it increases prompt complexity and token cost, and it does not guarantee deterministic execution for tasks that are fundamentally procedural.

Additional context
Implementing workflows as a distinct, structured resource allows for a clear separation of concerns. It enables client systems to use the appropriate execution strategy for a given task: deterministic logic for rigid processes and non-deterministic LLM reasoning for flexible, creative, or interpretive tasks. This architectural distinction improves the overall robustness and predictability of agentic systems built on top of the MCP server.

[remi-td]

If you really mean "exposing a resource that specifies a workflow", I believe that we can do this already with custom resources. Of course, this doesn't guarantee that the workflow will be correctly implemented by the LLM.

If the ask is to provide a workflow engine that executes this logic, the current alternative approaches are:

1. Specify the workflow using any programming language (eg. tools and ressources can be exposed as REST endpoints). Frameworks like LangChain can be useful.
2. Specify the workflow as a tool in the MCP server, possibly referencing other tools, using Python programming, that should be more appropriate to define complex workflows. The limitation is that we wouldn't be able to do "prompt chaining" type of patterns - if that's the requirement, the "client" tools would be better suited to define and execute this logic (ie. option 1).

[remi-td]

see alternatives above