tamara and Zachi score rows with Jev instead of prompting
Two builders replaced a generative step with per-item scoring, and Theo says the compaction version will make coding agents dumber.
Two builders arrived at the same shape this week. Instead of asking a model to write something, they ask Jev to score each candidate and keep or drop it.
tamara posted a compaction script she calls instant compaction, which scores every tool call in the history and drops what is irrelevant rather than summarizing it. Her framing is that in 2026 compaction is still a summarization prompt, and scoring removes the generative step. The script is published as fast-jev-compaction. Diogo Almeida replied approving of the direction, saying it would "free coding agents from designing around the KV cache".
free coding agents from designing around the KV cache
Zachi shipped the same idea in a database. His jev() Postgres extension turns plain English into a WHERE clause, so you write WHERE jev(people, 'could work from home') and every row is judged individually with no index and no embeddings. He measured 129 rows judged in about 1 second for $0.0009, and 6ms from cache on the second run. He said he was updating the live demo to around 4,500 rows to see how far it goes.
Theo says the compaction version is broken
Theo took tamara's script apart in detail. His first point is definitional, that compaction is meant to clean up history to keep the agent focused rather than delete noise, and should run sparingly rather than constantly. His second is that the scorer does not see enough to decide, since it works per tool call, knows little of what came before, and in that implementation does not even see the tool call result. He expects that to produce what he calls stupid loops, where the model retries things it already tried because the record is gone.
He also raises two costs that scoring does not fix. Frontier models from OpenAI, Anthropic, xAI and Google do not share reasoning traces over the API, they share encrypted payloads that Jev cannot see, and Anthropic requires the entire history be preserved to get reasoning data back. And editing history invalidates the cache from the edit point onward, so deleting item 2 from a six item history forces a rewrite of 3 through 6. Theo says cache writes are the biggest agent cost by far, over 60% of his own spend in Claude Code and Codex.
For people building on it, the two cases are not the same bet. Zachi is scoring rows in a database where nothing downstream depends on the ones he drops. tamara is scoring a conversation history where the dropped entries were the agent's memory, and that is the part Theo says will hurt you.
Compaction isn't a filter
129 rows judged in ~1s for $0.0009. Second run: 6ms from cache.