Convergent evidence

Multiple pieces of software were reviewed during a ten-day build, yet none of the nine tools were selected to be integrated. This demonstrates a strong argument for developing the software in-house. The process involved a companion project that monitored and assessed AI tooling for the entire system, ensuring that only tools adding unique capabilities to the orchestrator were considered.

Only a handful of candidates met this requirement, revealing a consistent pattern among the rejected tools. Six of the candidates were orchestrators or multi-agent systems solving the same problem of assigning tasks to different models and verifying their outputs. The remaining tools, including Ruflo, CrewAI skills, 9router, Claude Council, three agentic OS reels, and others, were rejected due to various reasons such as code vulnerabilities, overlapping functionalities, or lack of unique benefits.

Notably, a stranger's public routing configuration offered a fresh perspective by introducing "taste" as an additional factor beyond difficulty in model selection. This concept was deemed valuable enough to be incorporated into the project's routing policy. Overall, the convergence of multiple independent sources highlights that the same routing model was discovered by various parties, but none of the reviewed tools provided capabilities that the existing in-house system lacked.

The final conclusion is that building the software internally saves time and resources by avoiding dependencies on external solutions.