tkellogg ships skill so subagents pick models on real data

A small skill on GitHub from tkellogg caught my eye, not because the code is complicated but because of what it assumes about how agents ought to work. The pitch, in the author's own words, is "so subagents can better select models based on real data."

The skill queries Artificial Analysis' API for model performance metrics and organises them so an agent can pick a model based on cost and benchmark score for the specific task at hand. The prompt the author used to generate it asks for separate files per topic, one for coding benchmarks, another for writing, and so on, with the benchmarks' pros, cons and shortcomings written up so the agent has something to reason against rather than a single leaderboard number. Cost sits alongside performance, with the prompt explicitly asking for an example that shows cost per intelligence.

Why this matters if you are wiring up multi-agent workflows: today, model choice is usually a config decision made by a human at build time and then frozen. Push that decision into the agent itself and provider price cuts, a new release, or a bad day on a benchmark can propagate into routing without a code change. The install is deliberately small, a git clone into ~/.claude/skills, with an optional symlink into ~/.codex/skills for people running both harnesses off the same skill.

The honest caveat is that this is one person's quick build. The author openly says he "one-shotted this in Codex" from a single prompt, the repo has one commit on main, and the routing is only as good as the benchmarks Artificial Analysis publishes. API calls are cached daily to stay under the free tier's 100 requests per day, which is fine for a hobbyist but means a new model release will not show up in a routing decision for at least a day. The README also does not spell out how the skill actually enumerates the models available inside Claude Code or Codex at runtime, which is the load-bearing part of the whole idea.

Still, the pattern is what is worth watching. Skills that give an agent introspection about its own runtime, what models it can call, what they cost, what they are actually good at, are the plumbing that turns "the agent picks" from a slogan into an actual design choice.