Ship features with an AI team, not a single agent

A single agent asked to build a feature guesses the scope, writes tests that pass by construction, and reviews its own work. SoloDev splits those jobs across a team: you approve one spec, and you get back a reviewed, tested, documented pull request, on Claude Code, Codex, or Cursor.

Role
Creator and maintainer
License
Open source, Apache 2.0
Runs on
Claude Code, Codex, Cursor
In use
Daily, on every project

The story

  1. Problem

    “Use AI,” with no playbook

    ACT asked our team to adopt AI. I rolled out a customized Superpowers skill set. It worked, and it burned the monthly token budget in about two weeks.

  2. Rebuild

    A team, not one agent

    On my own time I rebuilt it lean: a commander and four role-separated subagents, a strong model to plan and review, a cheaper one to write code, and every rule a script can check moved into a CLI.

  3. Today

    One command, one ready PR

    /solo-task takes a feature from spec to a reviewed, tested pull request with QA results and a file-by-file walkthrough. I use it every day on every project, so my time goes to the problem, not the busywork.

One agent vs. a team

Ask one agent to build a feature and it fails in three predictable places. SoloDev gives each of those jobs to someone else.

Scope

Guesses what you meant and starts coding.

You approve a spec before any code exists. A CLI check fails any branch with code written before approval.

Tests

Writes tests that pass by construction. They were never seen to fail.

Each test is shown to fail first. The reviewer reads that failing run before running it on the new code.

Review

Reviews its own work, with the same context and the same blind spots.

The reviewer never wrote the code. Auth, payments, migrations, or a public API get one reviewer per lens, in parallel, and can run on another vendor’s model.

You get

A diff, and you start reviewing.

A ready pull request with QA results and a file-by-file walkthrough. It never merges or pushes to main.

Who does what, and what it costs

Each lane is a cost tier, and the CLI’s checks spend no tokens. Every step has a budget: 2 attempts, 1 fix round, 1 re-review.

YoudecideThinkstrong modelImplementcheap modelCLI0 tokens/solo-taska featureSpecscenariosApprovethe speccheckno code yetCodefail firstReviewfresh eyesDraft PRQA, readyMergeyour callfix onceYoudecideThinkstrong modelImplementcheap modelCLI0 tokens/solo-taska featureSpecscenariosApprovethe speccheckno code yetCodefail firstReviewfresh eyesDraft PRQA, readyMergeyour callfix once

Calls I made

  • You approve a spec before any code exists

    The spec holds requirements, tagged acceptance scenarios, and a plan. “No code before approval” is a unit-tested git check in the solodev CLI that fails the branch, not a line in a prompt that a model can skip.

  • Tests must fail before they pass

    Each failing run is recorded. The reviewer reads that record before running the test against the new code, so a test that passes by construction gets caught.

  • The reviewer never wrote the code

    Coder, spec reviewer, code reviewer, and verifier each run in fresh context on their own model, and the coder can run on a different tool, so one vendor’s model checks another’s work. Changes to auth, payments, migrations, or a public API get one reviewer per lens, in parallel.

  • The strong model thinks, the cheap model types

    Planning and review need judgment. Writing code to an approved spec mostly doesn’t. That split is where most of the savings come from.

  • Keep one coder warm in the cache

    One coder works the whole feature and is continued, not restarted, for each journey and fix round. Its cache lasts 1 hour instead of 5 minutes, so its context is reread at the cached rate. The reviewers start fresh and share the default cache.

How it’s built

Agents
  • A commander
  • One writer, report-only reviewers
  • Hard permission denies
Gates
  • Checks over git, not prompts
  • Unit-tested CLI
  • Retry budgets that stop and ask
Build
  • One source catalog
  • Native agents for 3 tools
  • Models pinned per role
Evals
  • 77 decision scenarios
  • 40 routing queries
  • Headless stage evals graded on git
Release
  • TypeScript CLI
  • Skill line budgets in CI
  • Migration paths across majors
Repo
  • Specs live outside it
  • State derived from git
  • Never merges or pushes to main

Try it

$ npx skills add luisintosh/solodev
Then, once in each coding tool, and in a new session:

/solo-task init
/solo-task Add CSV export to the accounts page
View SoloDev on GitHub
Next case · UltraA crypto wallet that unlocks like a normal sign-in