Tools, Agents, and Workflow Automation

Overview


In the previous part, we built command-line applications that send prompts to a model and handle the response. In this part, we go one step further: the model can now be connected to tools that let it retrieve data or trigger deterministic operations.

This is the basic idea behind many current “agentic” systems. The model does not only generate text. It can also decide that a tool should be used, provide parameters for that tool, receive the tool result, and continue from there.

That does not mean the model becomes an independent software engineer or a magical general-purpose agent. In useful systems, tool use is still wrapped inside software that defines the available tools, validates parameters, limits permissions, and decides when human oversight is needed.

Figure 1 summarizes the high-level loop used in many tool-using systems.

Fig 1. — Tool-using LLM systems extend the ordinary prompt-response pattern with explicit tool calls and tool results.
Loading Exercise...

Capability and authority are different things

Part 4 adds new capabilities, but it does not automatically grant the system new authority. A model may be able to suggest a tool call, but the surrounding program still decides which tools exist, what permissions they have, what arguments are accepted, and whether a human needs to confirm the result.

That distinction is important throughout the rest of the course. Many discussions about agents become confusing when capability and authority are mixed together. A useful system can be quite capable while still operating inside very narrow boundaries.

Loading Exercise...

The structure of this part is as follows:

Finally, Recap and Feedback summarizes the part and prepares you for the next part on retrieval and evaluation.