Case 01 / 08 · Agent development environment
Atelier
I don't work alone. I built Atelier, the system that runs my team of AI agents, so you get a team's output with one senior engineer in charge.
The problem
One engineer can only type so fast. The usual answer is a bigger team, which brings meetings, handoffs and mistakes nobody owns. What clients actually want is both: the speed of a team and the care of one senior person who answers for every line.
How it works
Six steps,one system.
- 01Brief
- 02Plan
- 03Build
- 04Prove
- 05Route
- 06Watch
Relja · Atelier · one chat
Brief written. Two questions before I plan:
1. Which currencies at launch?
2. Do refunds need a second approval?
I talk to one orchestrator. It already knows every past project.
A client signs. I tell Atelier everything in one chat: what we agreed, deadlines, constraints. Atelier writes the brief and asks about anything missing.
Inside
What I built,layer by layer.
01
One orchestrator per project
I write to one chat. Atelier turns it into a plan, starts the right agents, groups them by role and pauses them when something needs me.
02
A team with real roles
Every agent has a role, rules and its own tools. The security agent cannot deploy, testers cannot change code, and each one works in its own copy of the repository.
03
Memory that never gets mixed up
Knowledge lives in Markdown in git, task state in a database with locks. Twelve agents can work at once without overwriting each other, and every project leaves reusable patterns for the next one.
Engineering decisions
01
Done means proven
A different agent, starting from a clean context, runs the real app and attaches the proof: test log, screenshot or recording. Without it, the task stays open.
02
The right model for every task
Claude plans, designs and reviews. Heavy backend work goes to Codex at high effort, small fixes to a fast model. Every task records model, cost and whether it passed first time, so the routing gets better with each project.
03
Wake me only when it matters
A local model watches every project around the clock. It fixes what it safely can, and calls my phone only for something critical. Anything that reaches a client or costs money still goes through me.
Scale of the codebase
12
agents working in parallel
1
chat to run a whole project
0
tasks closed without proof
24/7
on watch, calls me if critical
Designed and built alone: the native app, the agent runtime, the project memory and the orchestrator. I use it every day on client work.
Next case
Dispatch →