LegalOn cut its daily OpenAI Codex costs by 65% without slowing down development. The company did it by routing work intelligently across three internal agents: Astra, Sol, and Luna, each matched to task types by complexity and cost profile.
The mechanism matters more than the headline number. Rather than applying a single model uniformly, LegalOn built budget-aware orchestration that assigns agents based on what a task actually requires. That architectural decision is the real story here, and the original piece details how that routing logic was structured.
Teams running multi-agent Codex workflows at scale should read this for the specifics on agent role design and budget governance. The 65% figure is the outcome. The decisions behind it are what transfer to other organizations.
[READ ORIGINAL →]