Use case · Engineering

Three hundred call sites.
One afternoon.

The billing-v1 client was deprecated two quarters ago. Three hundred call sites still use it, because no engineer has a spare month. A swarm has sixteen spare hours — all at once.

Migrations don't stall because they are hard. They stall because they are three hundred small chores wearing one big ticket. Split the ticket across sixteen workers and the chore part disappears what remains is the dozen decisions that actually need an engineer.

9:30 AM

The ticket everyone routed around.

An engineer opens the migration ticket that has survived four sprint plannings, and closes it differently this time: one request, one worker per module, and a rubric that defines what “done” means. The migration guide rides along as context.

one request — the whole migration
POST /api/v1/swarms{  "objective": "Migrate every caller of billing-v1    to the v2 client",  "tasks": [    "Rewrite call sites in modules/invoices (31 sites)",    "Rewrite call sites in modules/payouts (24 sites)",    "…one worker per module — 16 modules, 300 sites"  ],  "evaluation": "Per diff: types pass, tests pass,    no behavior change against the fixture suite",  "aggregatorTask": "One branch per module; flag any    diff scoring below 0.9",  "budgetUsd": 10.00}

9:32 AM

Sixteen branches grow at once.

Each worker gets a sandboxed checkout, its own module, and nothing else. They rewrite call sites, run the module's tests, and adjust until green — simultaneously, in sixteen isolated sandboxes, never inside anyone's dev machine.

The $10 ceiling is a hard line across the run. A worker stuck in a test loop gets stopped at its share — a stalled module costs cents, not the afternoon.

one call · four workers · one answer

you
researchrunning
draftrunning
fact-checkrunning
pricingrunning
merged

10:20 AM

The judge reads every diff.

An evaluation scores every diff against the rubric: types pass, tests pass, behavior unchanged against the fixtures. 288 call sites clear the bar. Twelve don't — ambiguous retry semantics, a hand-rolled mock, one module where v2 has no equivalent endpoint.

The point isn't that machines review machines. It's that the human's attention is spent only where the judge lost confidence.

evaluation scores · 16 modules

modules/invoices0.97pass
modules/payouts0.95pass
modules/webhooks0.93pass
modules/refunds0.71flagged — retry semantics changed

2:40 PM

Nothing merges itself.

Opening pull requests against the repo is an external write, so every branch holds until the engineer approves it — policy, not etiquette. She reads the twelve flagged diffs closely, fixes two by hand, and releases the rest module by module.

By the end of the afternoon the deprecation is a merged set of PRs with an append-only record of who approved what, and when.

policy: external writes require approval

agent requests: send 240 emails via connector
held for approval — nothing sent
executed · logged to the audit trail
The ticket was scoped at an engineer-month — call it $15,000 of salary nobody could spare. Sixteen workers rewrote 300 call sites in 50 minutes for $6.80, and the engineer spent one afternoon on the twelve diffs that deserved an engineer.
The uncomfortable math

0

call sites migrated in parallel

0

diffs flagged for human review

$0.00

total, against a $10 hard cap

The same shape, everywhere

Anywhere the backlog is many small diffs, a swarm wins.

Dependency upgrades

One worker per package bump — each with its changelog read and its test suite run.

Test backfill

Uncovered modules get a worker each; evaluations reject tests that merely assert the mock.

Lint-debt paydown

Ten thousand warnings sharded by directory and cleared without a human touching one.

Dead-code sweeps

Every suspected-unused export traced across the repo before a removal PR is proposed.

Incident repro hunts

Sixteen hypotheses about a flaky failure tested in parallel sandboxes overnight.

First-pass PR review

Each open PR gets a worker: style, tests, and risky-diff flags before a human looks.

Keep exploring

Give your agent a workforce.

Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.