# The Orchestration Receipt

> I pointed Claude Code's own fleet-orchestration primitive at Claude Code's own evolution. One hundred and thirty-seven agents across three runs, model-tiered, adversarially verified, and self-corrected. Here is the receipt.

Experiment 01 · 2026-06-09 · dispatch

Field-test dispatch from Ryan Hunter's Experiments.

This site sells the missing layers between AI agents and real organizations. One of those layers is workflows: the ability to take a single intent and fan it out across many coordinated agents, then bring the results back as one answer you can trust. This is the receipt for that layer.

The setup is recursive on purpose. Claude Code shipped a primitive called dynamic workflows, where one prompt orchestrates tens to hundreds of background agents. I used that primitive to study the primitive: I pointed it at Claude Code's own version history and had it map how the tool became a multi-agent orchestration platform in the first place.

## What ran

Three runs, each pushing the fan-out harder.

The first run was three agents in a line: extract, write, verify. The second fanned six versions through that pipeline in parallel and added a synthesis step at the end. The third went wide, fifty-six versions at once, one hundred and fifteen agents, and added a cross-version reconciliation pass.

One hundred and thirty-seven agents in total. Around three and a quarter million tokens. Every version documented, every claim traced back to the source changelog, nothing invented.

## The lesson that paid for itself

The third run is where the economics showed up. The first two runs used the most capable model for every agent, including the agents whose entire job was copying changelog bullets into a table. That is the equivalent of hiring surgeons to alphabetize a shelf.

So the third run was tiered. Cheap, fast models did the grunt work. The expensive model was reserved for exactly one seat: the cross-version reconciliation pass, the only step that needs the whole picture in one head at once. The result did six times the work of the first run while the premium model touched only a single node. Fan-out cost is a model-tier decision, not an agent-count decision.

## Honest by construction

A fleet of agents each writing on its own is a fleet of agents each tempted to look productive. Two guards held that in check.

Every finding was verified by an independent agent that re-read the artifact on disk and checked it against the source, rather than trusting the writer's word. And a final reconciliation pass held all the version records at once and asked a question no single-version agent could answer: was anything called new here that actually shipped earlier?

It found two. A feature attributed to one version had shipped three releases prior. A status field claimed as new had landed the version before. Both were corrected and both are documented in the record. The interesting part is structural: a fan-out of isolated agents is blind to that class of error by construction, so the system needs a deliberate step that sees across all of them. That step is where the expensive model earns its seat.

## The meta layer

The three workflows are not just runs that happened and ended. Each one is captured as a typed, content-addressed definition in the Workflow Graph Language, the canonical schema for durable multi-agent workflows. A fourth definition wires the three together as a single escalation, each step calling the next by content hash. The workflow that produced the corpus is itself a first-class object in the corpus.

That is the whole point. A workflow you can define, hash, replay, and audit is the difference between a clever one-off and a layer you can sell. The map of the territory and the vehicle that drew it are both on the table.

---
Source: https://ryanhunter.io/experiments/the-orchestration-receipt
From Experiments by Ryan Hunter. https://ryanhunter.io/experiments