Shipping dev-agents: an AI team in your editor

For the last month, I've been shipping dev-agents — a framework that turns Claude Code into something closer to a full engineering team. Not a single chatbot that tries to be good at everything. Twenty-one specialists: database engineers, security reviewers, frontend architects, testing experts, each with their own scope, rules, and decision-making style.

The problem it solves

Claude Code is powerful, but prompt length is real. If you ask it to build auth and design a schema and wire up the frontend and write tests — in a single message — you get either a 10,000-token response or the model cutting corners. You don't get a database expert thinking about your schema. You get a generalist trying to context-switch.

What if you could type this instead?

/lead Build user auth with PostgreSQL + a React frontend

Then watch five specialists work in parallel — each one deep in their domain — and come back with a proportional, coordinated response.

How it works

The framework has three layers:

• Slash commands — The entry points. /lead is the orchestrator. /commit, /review, /debug are quick triggers for common workflows.
• Specialist agents — The real work. Frontend, backend, database, security, testing, performance, infrastructure, incident response. Twenty-one in total, each with its own YAML frontmatter, tools, and scope.
• Shared rules — lib/universal-rules.md. One source of truth. Every agent reads it. No duplication, no drift.

When you call /lead, it clarifies your request, breaks it into a plan, then dispatches specialists in parallel using the Task tool. Each one gets just enough context to be expert-level, and they report back independently.

Or you can skip the orchestrator. Just describe a task in natural language and the right specialist auto-dispatches by description match.

What I'm using it for

Every major project I touch now goes through the agents framework. It's not because I've automated away my own thinking — it's the opposite. By handing off the routine tasks to specialists, I get to stay in the design problem. I ask /lead to implement a feature end-to-end, and what I get back is: correct schema, typed API contracts, tested handlers, accessible React, security checks, and a postmortem plan if anything breaks in production.

The time I save isn't hours. It's a different kind of time — the time to think clearly about what I'm building instead of context-switching between languages, frameworks, and domains.

Who this is for

If you're a solo engineer, a small team, or someone who pairs with Claude Code:

• You work across the full stack and don't have time to be expert in every domain.
• You want the same decision-making consistency that a senior team would give you.
• You're tired of the back-and-forth: "Oh, also make sure the password hashing is this way", "Don't forget the database index", "Add a11y labels to the form".

Then you probably want this.

What's in the box

The framework is open. You can install it per-user, per-project, or symlink it so edits are live. It's designed to be forked — if your team has different rules or specialists you want to add, the structure makes that obvious. YAML frontmatter. Clear file layout. No magic.

The idea is simple: be expert at what you do. Let the specialists handle the rest.

dev-agents is on GitHub. Install it today. If you find yourself rebuilding the same prompts or coaching Claude on the same principles over and over, you've probably already written half the framework in your own head.