Give your AI agent a task. Let it find the tools.
You have a task. Find a public transport route from Shinjuku Station to Haneda Airport Terminal 3 for a traveler leaving at 8 a.m. Tokyo time. Return the lines, transfers, walking segments, and a source the traveler can check.
A maps API could help. Before choosing one, work out which source covers the trip, whether it supports public transport, and whether you can access it with the credentials you have. Those decisions come from the task.
Choose a way to complete the task
Define a usable result. A list of nearby stations is insufficient. You need a route that starts at the right station, ends at the right terminal, and shows when and where to transfer.
Check coverage and access. The Japan pilot points to NAVITIME, which documents transit and walking route search through a business API. Check your access before planning around it. If you lack access, find a public journey planner and cite the source you actually used.
Verify the result. Use the travel date and local time. Cite the route source and flag any part of the itinerary you could not verify.
The same approach can lead to a simpler tool
If the task is to get Singapore’s official 24-hour forecast, the Singapore task page points to a public government JSON endpoint. Read the forecast and its validity window, then cite the source.
Task pages bring together websites, APIs, MCP servers and direct actions where available. The pilot identifies sources and their documented access requirements. We’re testing whether that helps agents reach a verifiable answer with fewer calls than ordinary search.
Try the Japan task
Open the pilot page. Find a route from Shinjuku Station to Haneda Airport Terminal 3 for tomorrow at 8 a.m. Tokyo time. Return the local date, train lines, transfers, walking segments and a source URL. If your first source is unavailable, name the fallback you used. Stop after returning the itinerary.