Orchestrate subagents at scale with dynamic workflows

Why it matters

When a task outgrows one context window, what matters is who holds the plan. A dynamic workflow puts it in a script, so verification can’t be skipped.

What to practise

From the Python Agent SDK, add the Workflow tool to allowed_tools and ask in plain words for a workflow that fact-checks a draft investor update against four source files. The prompt sets the structure: extract the claims, verify each in parallel, have a skeptic challenge every confirmed verdict, and compile a go/no-go report. Then read the JavaScript script Claude wrote: agent() calls with a schema for structured verdicts, parallel() as a barrier, and plain code between phases. Note the cost the run prints.

How it fits

The per-claim fan-out is 1.5.3’s partitioning: each verifier reads the shared sources fresh in a clean context. The skeptic pass is a recombination gate of the kind 1.5.4 describes, enforced by control flow. For 1.5.2, compare a plan held in a script with one a lead agent adapts turn by turn.