Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
This article explains how to break down agent flows into three core stages: input, transform, and output. You learn how to make informed decisions about whether to implement each stage by using unstructured (AI enhanced) or deterministic (code based) approaches.
In this article, you learn how to:
- Organize agent workflows into input, transform, and output stages.
- Choose between unstructured and deterministic components for each stage.
- Align your requirements with desired capabilities.
- Design effective solutions that balance flexibility with predictability.
This guidance helps architects and developers build robust agents by explaining component trade-offs and guiding technology choices for each workflow stage.
Organize agent workflows into input, transform, and output stages
You can organize agent workflows into three stages: input, transform, and output. Each stage can use either unstructured (AI-enhanced) components or deterministic (code-based) components, as shown in the following table:
| Stage | Dynamic flow (unstructured) | Deterministic flow (structured) |
|---|---|---|
| Input | "I'm marketing an automotive brand, get images of cars for use in my ad campaign." | Get top 100 work items sorted by last modified date. |
| Transform | Several tools could be used (image search, image generation), with unstructured parameters. | Query string is generated with parameters top=x, sortby = y for every request. |
| Output | Different images are generated on each run even if backend data doesn't change. | Must always get 100 results, with correct date and sorting order. |
Some components work best as deterministic solutions, like a user's ability to select their client of choice. Other components, like tools, might work best as dynamic or deterministic solutions based on the desired output and tolerance for variance.
Align workflow stages to technology
Mapping workflow stages to technology choices requires careful consideration of business requirements and operational constraints. The following table maps input, transform, and output requirements to agent options:
| Stage | Unstructured options | Deterministic options |
|---|---|---|
| Input | Knowledge bases, semantic indexes, images, agent-to-agent communication | API Plugins, Power Platform connectors, Model Context Protocol (MCP) tools |
| Transform | Generative AI models | Power Automate flows, Logic Apps, code-based logic |
| Output | Generative AI responses including declarative agents | Adaptive Cards, topics, MCP tools |
Note
- Transform and output stages can only act on data that the agent receives as input. You need to understand the formatting and size requirements of inputs including links, table headers, row counts, images, and full document context to generate desired outputs.
- The principle of "garbage in, garbage out" applies to agent implementations. Truncated or chunked content from semantic indexes missing required metadata degrades output quality regardless of the sophistication of transform and output components.
Assess capabilities
Each approach provides different levels of support for unstructured and deterministic approaches across workflow stages.
| Stage | Unstructured options | Deterministic options |
|---|---|---|
| Input | Declarative agents via semantic index, Custom engine agents via semantic index | Custom engine agents, declarative agents via plugins or MCP |
| Transform | Declarative agents, custom engine agents | Declarative agents via plugins or MCP, custom engine agents |
| Output | Declarative agents, custom engine agents | Custom engine agents |
Use this capability matrix to help you decide which approach to take based on the specific combination of unstructured and deterministic requirements for each workflow stage.
Learn more:
Next step
Learn how to evaluate requirements that guide technology choices for each agent component.