Building agents

An agent runs a controlled loop: it sends messages and tools to the model, receives a function request, validates and runs the action, returns the result, and repeats until it reaches a final answer.

1. 1

   Declare small tools

   Use specific names, objective descriptions and parameters constrained by JSON Schema.

2. 2

   Validate before running

   Treat arguments as untrusted input. Check types, permissions and limits.

3. 3

   Return the result

   Include the tool message with the matching tool_call_id.

4. 4

   Enforce limits

   Set a maximum number of steps, a timeout, retries and a budget.

5. 5

   Confirm high-impact actions

   Require human approval before charging, deleting, publishing or sending.

Choosing the model

Start with HiNova for general use. Consider HiMax when the workflow requires planning, many constraints or complex analysis; consider HiGenesis for simple, repeatable, high-volume tools. The final decision should come from your own tests.

Production checklist

• Log calls, tools, arguments, results, cost and duration.
• Use idempotency for repeatable actions.
• Return structured errors so the model can correct course.
• Block data and tools outside the user's scope.
• Test success, refusal, timeout, retry and empty results.
• Provide a safe exit for the step limit.