The Mechanics of Agency

• The language model itself is strictly a text generator; it cannot call tools, save memory, or execute code. Software orchestration must capture the model's text output and execute functions on its behalf.1:17
• Memory is not intrinsic to LLMs. To maintain conversation history, developers must manually append prior system, user, and assistant messages into a list and re-send that full history with every request.7:24
• Tool use requires a structured schema describing functions, which the model references to request specific actions. The software harness must then parse these requests, execute the associated local Python function, and feed the result back to the model.17:23
• An agent operates as a loop that repeatedly prompts the model to assess its next move, processes requested tool calls, and appends outputs until the task is complete.1:52

Implementation and Safety

• A basic coding agent can be built in under 200 lines of pure Python, demonstrating the capability to list files, read code, write files, and execute shell commands to build projects like a playable pygame snake game.33:24
• Destructive tools like shell execution require a human-in-the-loop guardrail. The demo implements a mandatory user-approval step before executing any terminal command to prevent unintended file system modifications.19:13
• HubSpot promotes a four-part decision tree to qualify tasks for automation, suggesting businesses should only automate processes that are repetitive, use structured data, accept 90% accuracy, and have measurable outcomes.12:22