Apply task decomposition and agent collaboration strategies in Microsoft Foundry

Why it matters

Decomposition is where an agentic design decides what a model may plan and what code must enforce, and this exercise makes that split concrete.

What to practise

Deploy the Foundry resources with Bicep and azd, then complete a planner agent that returns a JSON task DAG. Your code validates the plan, invokes specialists only when their dependencies are complete, passes size-bounded handoff envelopes, and raises at most one evidence-driven replan signal for an outer coordinator. Compare a simple and a complex query: the captured task, handoff and replan counts and the coordination ratio show what the extra plan depth costs, and the Foundry traces show whether dependent specialists start only after their prerequisites finish.

How it fits

The validated DAG is 1.1.1’s bounded work and 1.1.2’s deterministic control around a model’s plan; the handoff envelope is a delegation contract from 1.1.3; the single replan signal is 1.1.5’s enforced termination. Ask which plan invariants you would never leave to the model.