← Table of Contents · Page 156 of 868

DESIGNING GOOD TOOLS

Part 3 of 10

MINIMAL AND COMPOSABLE Composable means able to be combined. Think of LEGO bricks: small, simple pieces that can be snapped together into many different creations. A giant tool called manage_project that takes a mode setting and fifteen optional inputs to create, assign, update, delete, and report on tasks is hard for a model to use correctly. Splitting it into create_task, assign_task, and mark_task_done gives each tool one clear purpose and a short input list. Benefits of small tools include clearer names, simpler schemas, easier testing, better safety (each action tool can have its own permission), and flexible combinations for tasks you did not predict. There is a balance, though. Too many micro-tools force the agent to chain a dozen calls for a simple job, costing time and money. Sometimes it is wise to offer a convenience tool that combines a common sequence, as long as it still has one clear purpose. A useful guideline is to design tools around the tasks the agent actually performs, not around the internal structure of your database. Another good habit is to limit what a tool returns. If a search tool dumps 50 pages of text into the context, the model drowns. Return the most useful fields, support paging (getting results in small pages), and let the agent ask for more detail with a follow-up tool.
Advertisement
Read Aloud:
0:00 / 0:00