Three tools with basic verb-noun naming (get_weather, get_market_price, check_pesticide_safety). All have descriptions (60-120 chars, within acceptable range). Input schemas present with type and description for each parameter. However, output schemas are NOT documented, LLMs cannot predict return structure. No error handling guidance (e.g., 'If district not found, try search_districts'). No parameter constraints (enums, ranges). Tool descriptions lack WHEN/WHY context. No tool annotations (readOnlyHint, idempotentHint). Composition is sound (each tool does one thing), but descriptions are minimal and output contracts are invisible.

• Document output schemas for all three tools. Example for get_weather: 'Returns {district: string, current_conditions: {temperature_celsius: number, relative_humidity_percent: number, precipitation_mm: number}, forecast: [{date: string, temp_max: number, temp_min: number, precip_sum: number}]}'.
• Add error recovery guidance to descriptions. E.g., get_weather: 'If district not found, call with a major district name (Dhaka, Rangpur, Chittagong, etc.) or ask the user to clarify.'
• Replace free-form string parameters with enums or pattern constraints. For get_weather, provide a list of valid districts in the parameter description or as an enum. For get_market_price, list common crops (Rice, Wheat, Maize, etc.).
• Expand tool descriptions to 100-150 chars with WHEN/WHY context. E.g., get_weather: 'Fetch current temperature, humidity, and 3-day forecast for a Bangladeshi district. Use this to advise farmers on irrigation, pest management, or harvest timing.'
• Add tool annotations: set readOnlyHint=true and idempotentHint=true for all three tools. This signals to agents that these calls are safe to retry and cache.