Principles of Building AI Agents

Suggested reading: “o1 isn’t a chat model” by Ben Hylak

The key to effective tool use is clear communication with the model about what each tool does and when to use it.

in Mastra we define the conditional path execution on the child step rather than the parent step. In this example, a processData step is executing, conditional on the fetchData step succeeding.

At a low level, agents make binary choices in a decision tree At a medium level, agents have memory, call tools, and retry failed tasks At a high level, agents do planning, divide tasks into subtasks, and manage their task queue.

Think of agents as AI employees rather than contractors: they maintain context, have specific roles, and can use tools to accomplish tasks.

You should describe both what it does and when to call it.

Best practices: Provide detailed descriptions in the tool definition and system prompt Use specific input/output schemas Use semantic naming that matches the tool's function (eg multiplyNumbers instead of doStuff)