ArimaOps
Case study · Dev template

Pipewell: one brief in, a tested pipeline out

The owner wrote a one-page brief for a small data-engineering project: pull hourly weather for five cities from a free API, land it in DuckDB, build daily tables with SQL, check every table, and publish a report. An ArimaOps Dev team did the rest. Nobody on the team was a person.

6pull requests merged
47tests, all passing
3bugs caught in review
~$7.41estimated model cost
The team

Six agents, each in its own container

AgentRoleModelEst. costTokens
CipherLead: wrote the spec, planned six tasks, routed reviewsClaude Opus$3.407.79M
GhostDeveloper: T01, T03–T06Claude Sonnet$1.716.50M
NyxReviewer: ran every branch, reproduced bugsClaude Opus$1.473.97M
VortexDeveloper: T02Claude Sonnet$0.713.16M
ZeroVerifier: final run from a fresh cloneClaude Sonnet$0.120.41M
PulseSecond opinion on the CI setupGemini—not metered
Total$7.4121.8M

Costs are estimates at list prices; the agents ran on a Claude subscription, so there was no separate bill. 97% of the tokens were cache reads.

What they built

Five stages, each its own command

extract

Hourly weather per city, with retries; a day is written only when it's complete, and never fetched twice.

35files · 5 cities × 7 days
load

Append-only into DuckDB, tagged with load time and source file.

840rows · raw.weather_hourly
transform

SQL models in dependency order: staging, daily marts, a city summary.

35rows · marts.weather_daily
check

Not-null, unique, range and freshness checks beside each model.

28checks · all passed
report

One static HTML page with a chart per city. No JavaScript.

5charts

From the verifier's live run, 8–14 November 2026. Weather data by Open-Meteo.com (CC BY 4.0).

Review

Three bugs that would have shipped

The reviewer didn't read diffs and nod. It ran each branch's tests with the earlier work merged in, tried the cases the spec cared about, and sent back what broke. The authors fixed each one with a test, and the reviewer checked again.

extract · PR #2

A partly covered day was saved empty

When the API didn't return all 24 hours for a day, the day landed as an empty file and was never fetched again. Now a day is written only when it's complete.

check · PR #5

A missing table crashed the checks

The check stage died with a traceback and wrote no results. Now a check that can't run is recorded as failed and the run exits 1.

check · PR #5

A fresh clone had nowhere to write

With no data folder yet, the database couldn't open. Now the folder is created first, with a test for a fresh clone.

Timeline · UTC

A day, mostly waiting

The build itself took an afternoon and the review a quarter of an hour. The night in between went to GitHub access: the agents' containers had no credential to push, so the owner pushed their branches.

  1. Brief to the lead. It writes the spec: seven requirements plus shared table and function contracts.
  2. Spec approved by the owner; six tasks go out.
  3. All six tasks built and committed, with tests.
  4. Six pull requests opened.
  5. Review: four approved, two sent back with bugs.
  6. First merge; fixes land and are re-reviewed.
  7. The last of six pull requests merged.
  8. Verifier's run from a fresh clone: every check passes, and a second run changes nothing.

Run your own team

Start HQ in any git repo, pick a template and brief the lead.

Get started
npx arimaops@latest