Orchestrator pattern: big models for planning, small models for execution

Why it matters

A cost claim for an agent design means something only against a baseline that delivers the same outcome. This exercise builds that comparison and reads the real bill from the API.

What to practise

In the Claude Managed Agents beta, define a worker scoped to web_search and web_fetch, and a frontier orchestrator with no tools, only a multiagent block that lists the worker and disables inline agents and workflows. Give the session a budget and ask a twenty-fact coverage question. Run the control: one frontier agent that must verify each fact from two independent fetches. Read each thread’s usage.list_cost, and price the team’s workload at frontier rates. Watch the share of input tokens billed at the worker rate, each run’s wall-clock time, and each worker thread’s fixed overhead.

How it fits

This is 1.6.2’s ledger at small scale: price the accepted outcome, not the cheapest call. The matched verification standard is 1.6.1’s comparable baseline, and the wall-clock comparison feeds the latency targets in 1.6.3.