Back to quick projects
The agent workstation
How we work every day: several AI models, each on the kind of task it does best, reviewing one another and never writing to production without approval.
Internal method: no code or configuration published.
Context
Coding agents are powerful but costly, uneven across tasks and prone to overconfidence. The question is not whether to use them, but how to frame them so they ship code you can put in production.
What was delivered
- An agent server reachable from a phone, rebuilt identically on every start
- Task-based routing between Claude, Codex, Grok and Gemini, adjusted to remaining quotas
- Cross-reviews: code written by one model is reviewed by another
- Written guardrails: no production write without approval, an evidence report before any risky merge
- Skills shared by every agent, versioned like code
Outcome
This is the method that built the other projects on this page.