Agents and tools

• Workflows are best when you know the exact sequence of steps needed
• Agents are best when the steps are unknown—you provide Claude with a task and tools, then Claude creates its own plan
• Agents offer flexibility: build once, solve many different tasks

• Agents combine tools in different ways to solve varied tasks
• Claude can determine which tools to use and in what order
• Claude can recognize when it needs additional information from the user before proceeding
• The same set of tools can be used in surprising and unexpected combinations

• Provide abstract, general-purpose tools rather than hyper-specialized ones
• Tools should be composable—Claude figures out how to piece them together
• Example: instead of a "refactor code" tool, provide file reading and bash execution tools
• Example: instead of "install dependencies," provide tools to read config files and run bash commands

• Bash (access to ffmpeg for video generation)
• Generate Image tool
• Text to Speech tool
• Post Media tool

These four tools enable Claude to handle both straightforward requests (create and post a video) and dynamic workflows (generate sample image → get user approval → proceed with video generation)

• Major drawbacks exist but will be discussed later in the course

Takeaways

• Agents work best when steps are unknown—Claude determines which tools to use and in what order, while workflows suit known sequences
• Design tools to be abstract and composable rather than specialized, letting Claude combine them in unexpected ways
• Agents offer flexibility to solve many different tasks with the same tool set, but require careful tool design to enable dynamic workflows