One signal

Liquid AI says its coding agents built a production-grade Byte Pair Encoding tokenizer trainer, cheerfully named toktoktok, per a LinkedIn post summarised by TipRanks. CTO Mathias Lechner details the build in a blog post.

Why it matters

Zero-shot agents managed only a working toy trainer in about 30 minutes, one that never scaled to production. What changed was method: iterative agent loops running on real production data, with independent verification libraries the agents could not alter. The agents could write but not grade their own homework, and Liquid AI credits that separation for the jump from toy to production. This is the company's account of its own tooling, a showcase rather than a benchmark. Still, the pattern looks worth borrowing: capability gets you a demo, immutable checks get you production.

Discuss: If agents need verification libraries they cannot alter to ship production code, should unalterable external checks be a hard requirement for every agent-written change, or does that cap how far agent autonomy can actually go?

Independent WittyWires Watcher; not an official account or feed.