---
source_hash: 69c84289607d4e0141a3a4aa126bbc2938a6799cdd3d6a0c299084c1c3dd72e9
post: /posts/workflows/swap-the-orchestra
shape: doctrinal
---

# Context Pack: Swap the orchestra

Agent reference for the rewritten June 2026 run report. Only the quoted excerpts are verbatim. Historical numbers and provider choices are not current price quotes or a guarantee of compatibility.

## Verbatim excerpts

> In the run behind this June 2026 essay, I dispatched 235 agents to review a codebase.

> A cheaper call repeated hundreds of times is still hundreds of calls.

> Parallel reviewers can catch each other's mistakes, but agreement between them doesn't establish that they're right.

## Argument and evidence

Separate the harness's tools, dispatch and coordination from the model supplying each seat's responses. The reported code-review run retained fan-out, adversarial verification and synthesis while using a budget model under Claude Code.

The numbers were 235 agents, 13.12 million tokens and an estimated cost of roughly one tenth of a premium-rate comparison. The original fifty-for-one phrase was an illustration of the breadth argument, not this measured cost ratio. Retry and review costs can change the economics. Ari still reviewed the results.

The cheaper worker in this setup offered less depth per agent. More attempts and independent review helped make broad exploration useful, but agreement does not prove correctness. Check the findings and remaining human correction, not only spend or agent count.

## Product detail and compatibility

Claude Code dynamic workflows generate orchestration scripts. The `ultracode` setting combines xhigh effort with automatic workflow decisions, as documented in [Anthropic's introduction](https://claude.com/resources/articles/introducing-dynamic-workflows-in-claude-code). This is not a guarantee that every request spawns hundreds of agents.

GLM and MiniMax are alternatives discussed in Ari's historical setup. Provider substitution requires compatible tool and workflow behaviour; the essay does not establish universal official support for every provider. Test a bounded task with unchanged acceptance criteria before broadening use.

## Retained illustration and references

`apps/web/components/blocks/ProviderCostDial.tsx` uses a historical June 2026 table and an assumed 45,000 input plus 11,000 output tokens per agent. It is not a live price feed. `WorkflowAnatomyAnim` and `SwapOrchestraAnim` show the coordination pattern.

Canonical body: `content/posts/workflows/swap-the-orchestra.mdx`. Related migration account: [Port the site, keep the conductor](/documents/dispatch/the-conductor-never-moved).
