Liquid AI published an engineering account on 18 August of an experiment it ran in late 2025: could coding agents autonomously solve a production-grade problem from scratch? Two agents using the best publicly available coding models at the time were given a real problem and a real deadline, and the output — a byte-pair encoding tokenizer trainer called toktoktok — is now open source.
The choice of target is the methodological part. The team needed a BPE tokenizer trainer able to process trillions of tokens on a single machine for its own research on vocabulary size in edge models, and found the existing landscape thin: sentencepiece was optimised for non-BPE tokenizers and slow. So the task was not a benchmark exercise but a gap they would otherwise have had to fill themselves, which is the only setting in which autonomy can be judged honestly. The write-up covers how to specify a goal for multi-domain experts and how to set up the verification infrastructure around the loop.
Verification infrastructure being the deliverable rather than the prompt is the pattern in agent work moving into harnesses, factories and accepted artifacts and the constraint named in verification bandwidth being the scarce resource. The agent harness hub collects the loop designs.
This is a single retrospective experiment on one problem with models from late 2025, and the account is the operator's own without independent replication. The condition to watch is whether the loop is re-run with current models against a second real gap, because a technique demonstrated once on a favourable target is a case study rather than a method.