Seven hours, $400 of tokens, 13,000 lines of Go↗ (external link)

Nir Barak at Reco had a JSONata problem: the reference implementation is JavaScript, their policy engine is not, and crossing that boundary cost about 150 microseconds per evaluation across billions of events — roughly $300k a year in compute to keep an RPC fleet alive purely as a language adapter.

So he had a model rewrite JSONata 2.x in Go. Seven hours. $400 in tokens. Thirteen thousand lines, passing all 1,778 test cases from the reference suite. They call it gnata, and it is in production.

Every instinct I have says to be sceptical of this post, and every specific thing in it survives the scepticism. Look at what they actually did before promoting it: code review, then preprod in shadow mode, then live only after three consecutive days of zero mismatches against the reference implementation while processing billions of real events.

That is the part to steal. This was not "the tests pass, ship it" — the tests passing was table stakes, and 1,778 of them is a luxury most projects rewriting something would kill for. The actual verification was a differential oracle running on production traffic, where the existing implementation votes on every result the new one produces. You cannot always get that. When you are reimplementing a specification that already has a reference implementation and a conformance suite, you can, and this is what it looks like done properly.

Which is also the limit of the story. This is the friendliest possible shape of problem for an agent: a closed, well-specified language with an executable definition of correct, and a free oracle sitting right there. Almost nothing in your codebase looks like that. The seven hours is not the transferable part.

On the money: $500k/year is $300k of RPC fleet plus about $200k from a later refactor that gnata made possible. The second number is doing a lot of work in the headline, and it is a projection of a projection. The $300k is the one I believe.