For those interested, feel free to take a look at my write up on the need for a standard routing protocol [0], or point your agent on the GitHub repo and let your agent install the role-model runtime. I suggest using Pi for now; I plan to also launch an extension/plugin for OpenCode. The runtime is an alpha release so expect bugs and such. I would welcome feedback on how to make configuration more intuitive for first-time users.
I will also add that based on my current experiments, the ideal number of models in a routing pool is probably 2, following point 4 above. Each model needs to have significant differences in either quality, speed, or cost, otherwise routing decisions are hard to make and become less accurate; the benefit of routing is also less. For coding, the ideal pool in my opinion is GPT 5.4 and DeepSeek V4 Pro, to extend the GPT quota by routing some of the medium and easy requests.
For those interested, feel free to take a look at my write up on the need for a standard routing protocol [0], or point your agent on the GitHub repo and let your agent install the role-model runtime. I suggest using Pi for now; I plan to also launch an extension/plugin for OpenCode. The runtime is an alpha release so expect bugs and such. I would welcome feedback on how to make configuration more intuitive for first-time users.
I will also add that based on my current experiments, the ideal number of models in a routing pool is probably 2, following point 4 above. Each model needs to have significant differences in either quality, speed, or cost, otherwise routing decisions are hard to make and become less accurate; the benefit of routing is also less. For coding, the ideal pool in my opinion is GPT 5.4 and DeepSeek V4 Pro, to extend the GPT quota by routing some of the medium and easy requests.
[0] role-model - the case for a model routing protocol: https://try.works/role-model-the-case-for-a-model-routing-pr...
Interesting, any benchmarks for how well it works?