The arc, and the workflow that drew it.
Two pictures. The escalation arc: how Claude Code became an orchestration platform, one layer at a time, in release order. And the four WGL definitions that ran this experiment, each drawn as what it actually is, a layered dependency graph, so you can read the fan-out width and the model tiers straight off the page. Box weight is the model: heavy gold is Opus at a barrier, light grey is a Haiku worker. The whole economics of the run is in that contrast.
The agentic escalation arc
Seven keystone releases, in ship order. Color is the capability layer each one added. ★ marks the keystone: dynamic workflows, where the platform learned to author its own fan-out.
The four WGL workflow definitions
Each ran a real slice of this experiment. Read them left to right: a node only starts once its inputs land. Watch the fan-out grow 1 → 6 → 56 versions, and watch the model tier drop to Haiku for the wide cheap work and climb back to Opus only at the narrow expensive barrier. Dashed edges are call (a workflow invoking another by content hash); solid edges are data dependencies.
wgldef_31f88f79f070957… Analyze one Claude Code version (2.1.154) as a sequential extract -> write -> verify chain. 3 agents, all Opus.
wgldef_1a51229e8026ece… Fan out across 6 versions (extract -> write -> verify per version), then a synthesis barrier writes the cross-version arc. 19 agents, all Opus.
wgldef_3b8976cf3aea066… Full flow across 56 versions. Haiku workers, Sonnet for the 6 meatiest, Opus only at the cross-version reconcile + synthesis barriers. 115 agents.
wgldef_a86910d24df2697… The single -> multi -> fullflow escalation as a meta-workflow: each step calls a concrete ultracode def by content hash.