Should the weather service have separate geocoding and forecast tools / skills?

[MrCsabaToth]

OpenMeteo based weather agent is a common example service in my toy multi agent systems. Usually it is used by a ReAct agent and I prefer to have the geocoding and the weather forecast tools / skills exposed separately. When they are separate, the forecast requires a lat/lon, so first the geocoding needs to be called, and that will yield the lat/lon for the forecast call. I noticed when I code such an agent with AI, it tends to create such a code which "embeds" the geocoding code under the forecast, which is also what I see here:

coffeeAgntcy/coffeeAGNTCY/coffee_agents/lungo/agents/mcp_servers/weather_service.py

Line 65 in ceab2c5

coords = await geocode_location(client, location)

IMHO it helps debugging when there are two separate tools / skills and the agent which uses the service (probably via MCP) will be intelligent enough to first call the geocoding and then the forecast with the lat/lon. I've found examples which are quirks of OpenMeteo's geocoding where specifying a US state throws off the call. The "tool user" AI agent probably tries to help and be more specific, but in some cases that fails:

1. Fresno, CA: it fails in this format, or also as "Fresno, California". Whereas geocoding "Fresno" (without the state) succeeds. Note, that Fresno as a city can be ambiguous, so I think the agent is right and the geocoding API is weird.
2. Same with Nashville, TN: it fails in this format, or also as "Nashville, Tennessee". Whereas geocoding "Nashville" without the state succeeds. Note, that Nashville as a city (just as probably every city in the US) can be ambiguous (for example Nashville, Arkansas).

When it fails like that the ReAct agent can readjust and call the Geocoding with other variations, and so the separated forecast won't pose an extra layer between the geocoding retries (when this happens).

[MrCsabaToth]

I didn't want to create an issue right away, because this is a minute detail, and having only one tool reduces the tool call count and could work better for SMLs, however as I mentioned when there's a geocoding issue that could be a different story.