Add a bounded workflow, then consider an agent

A fixed workflow runs steps you define. An agent lets a model choose some of its next actions. Choose the least complex approach that passes your task evaluation. Anthropic’s engineering guidance also recommends beginning with simple patterns and adding complexity when it is justified. Read the engineering article ↗.

Practice project: support assistant with two tools

1. Provide a read-only lookup tool for a synthetic product catalog. Define an input schema and deterministic output.
2. Provide a draft-ticket tool that prepares a proposed change. Require a human to approve any real submission.
3. Limit the workflow to a small number of calls and a timeout. Record tool names, validated arguments and outcomes.
4. Handle missing records, invalid arguments, duplicate requests and tool errors. Enforce permissions outside the model.
5. Compare the tool-enabled version with your simpler retrieval app on the same tasks. If it is slower and no more useful, keep the simpler app.

When to learn a framework

After you have written and tested the direct version, try one framework if you need persistent state, branching, resumable steps or clear tracing. Learn the concepts before collecting framework names. Use the Hugging Face: AI Agents Course ↗ Unit 1 for a structured introduction; select one implementation later instead of completing every framework track.

A note on MCP

Model Context Protocol is one way to connect AI applications with tools and data. It is useful when a target role or your integration needs it. First understand the tool schema, authentication, authorization and failure behavior. Making a server discoverable is not the same as making its actions safe.