YepAPI/skills

AI Agents

Rules

• Define tools as typed functions with descriptions — the model picks which tool to call
• Tool schema: { name, description, parameters: zodSchema, execute: async (args) => result }
• Multi-step loop: send messages → model returns tool call → execute tool → feed result back → repeat until done
• maxSteps limit to prevent infinite loops — default 5-10 steps depending on complexity
• Structured outputs with Zod: generateObject({ schema: z.object({...}) }) for guaranteed shape
• Retry logic: wrap tool execution in try/catch, retry transient failures 2-3 times with exponential backoff
• Agent orchestration: break complex tasks into sub-agents with focused tool sets, coordinator agent delegates
• Always validate tool results before feeding back to the model — sanitize untrusted data
• Use streamText with onToolCall for real-time visibility into agent reasoning

Patterns

// Tool definition
const tools = {
  searchDatabase: tool({
    description: "Search the product database",
    parameters: z.object({ query: z.string(), limit: z.number().default(10) }),
    execute: async ({ query, limit }) => db.products.search(query, limit),
  }),
};

// Agent loop
const result = await generateText({
  model,
  tools,
  maxSteps: 10,
  messages,
});

Avoid

• Giving agents unrestricted tool access — scope tools to the task
• Missing maxSteps — agents can loop forever
• Trusting tool outputs without validation — always check before acting on results
• Monolithic agents — prefer focused sub-agents over one agent with 20+ tools