CHAPTER 02 / SOFTWARE SYSTEM
How does an AI capability become something people can use?
Applications, workflows, and agents
An application combines a model with an interface, information and rules for doing work. A workflow follows a planned sequence; an agent can let a model choose some of the next steps and tools.
More than a chat box
An AI application may classify images, recommend items, transcribe audio or help someone write. The interface collects input, software prepares a request and the application presents an output. APIs are defined interfaces through which software systems exchange requests and results.
Workflows and agents
In a workflow, developers prescribe the path: retrieve a document, summarize it, then ask for review. An agent can choose a tool or a next step based on its current context. This distinction is a useful design lens rather than a universal naming standard. More freedom also creates more behavior to test.
Tools need boundaries
A tool can search a database, read a file or change an external system. Permissions, input checks, logs and approval steps determine what it may do. Giving a model access to a tool does not make every proposed action appropriate.
Keep this connection in mind.
Look at the complete workflow: model, tools, data, permissions and human review.
- People uses AI applications
- AI applications can support Human tasks
- Model-based applications uses Models
- Retrieval-enabled applications can retrieve from Information collections
- Responsible organizations governs AI applications
Sources for this introduction
- Anthropic · Building effective agents ↗
A practical distinction between workflows and model-directed agents.
- NIST · AI Risk Management Framework ↗
Human context, evaluation, accountability and risk management.
Draft teaching text, prepared with AI assistance. Larry’s editorial review is pending. Read the editorial approach.
EXPLORE RELATED AREAS