The Agent Fleet
Four small, single-purpose internal agents — one each for weather, time, and currency conversion, plus an orchestrator that delegates to the other three — demonstrate a reusable template for adding a new narrow-purpose agent to this org without inventing a new shape each time.
Who this is for
Anyone building the next single-purpose agent in this org (a template already followed once more by the podcast pipeline, see Daily News Podcast), and anyone using the orchestrator agent itself to answer a combined question ("what’s the weather and local time in Tokyo?") without knowing or caring that it’s actually three separate services underneath.
Why a fleet of small agents, not one big one
Each small agent does exactly one thing. Keeping them small and single-purpose means each one is independently deployable and testable, a new capability is additive rather than a change to an existing agent, and the orchestrator composing them proves the pattern actually works end-to-end for a caller who wants one combined answer.
Delegation goes through a small internal router, reasoning goes straight out
The orchestrator delegates to the other three agents through one small internal router that bundles them behind a single in-cluster address — the caller doesn’t need to know or care that "weather," "time," and "currency" are really three separate processes underneath. This isn’t an AI gateway: it does no reasoning, no model routing, and no request tracing of its own — it’s plumbing, trusted purely because network policy only lets the orchestrator reach it in the first place, not because of any credential on the call. The orchestrator’s own reasoning — the part that decides how to combine three agents' answers into one — goes straight to the outside AI provider, the same way every other AI call in this fleet does. See LLM Gateway for that direct-access pattern.
Architecture
Full deployment shape lives in Namespaces & Apps and LLM Gateway.