Claude Code Workflow
Fan out subagents, run them through ordered pipelines, and verify every result before you trust it.
Workflow lets you script many Claude Code subagents with real control flow. You decide what runs in parallel, what runs in sequence, and what has to be proven true before it counts. The model handles the reasoning; your script keeps the run deterministic.
Core building blocks
Launch many independent agents at once across files, services, or sources, then collect every result into one place. Wide coverage in a single pass instead of one slow sequential crawl, so a thousand-item job finishes in the time of one.
Push each work item through ordered stages with no barrier between them, so a finished item moves on while the rest are still in flight.
An independent skeptic agent tries to refute each finding before it is accepted. Only claims that survive the challenge reach you.
Keep spawning finder agents until no new results come back, or until a token budget is hit. You set the stopping rule; the workflow respects it and reports where it stopped.
Write loops, conditionals, and stage order in a script. Determinism lives in your code, not in a prompt the model might drift away from.
The runner spawns one subagent per work item with its own fresh context, so a thousand-file task is never bound by a single window.
Each result passes to a skeptic agent that argues against it. Findings that cannot be defended are dropped before they reach your report.
Results stream back into one structured output. The loop continues until it runs dry or hits the budget you set, then hands you the summary.
Anything where the work outgrows a single conversation and the answer has to be right.
Fan an agent over every module, gather findings, and let a skeptic cut the false positives before a human ever opens the report.
Run the same transform across hundreds of files in parallel, with a verification stage confirming each change still compiles and behaves.
Loop finder agents across many sources until no new material surfaces, then reconcile the results into one cited synthesis.
Sweep the tree for risks, challenge every flag adversarially, and report only the issues that hold up under scrutiny.
Authored, not improvised
A workflow is a script. Loops and conditionals are explicit, so a run is reproducible and reviewable.
// Fan out one agent per module, then verify what comes back.
const modules = await listModules("./src");
const findings = await fanOut(modules, (m) => {
return agent.review(m, { focus: "security" });
});
// Adversarial pass: a skeptic must fail to refute each finding.
const verified = await verify(findings, {
refuteWith: skeptic,
keepIf: (f) => f.survivesChallenge,
});
// Loop until no new findings, capped at a token budget.
await loopUntilDry({ budget: 200_000 });
report(verified);