Theo tears down Jev instant compaction in six points
A cheap-model trick for shrinking agent context runs into how caching and reasoning traces actually work.
tamara (@tamarajtran) posted what she called the perfect use case for Jev, the cheap classifier model, which is instant compaction. Instead of summarizing a long agent thread with a prompt, score every tool call and drop whatever is irrelevant. Her framing was that in 2026 compaction should not still be a summarization prompt.
Theo answered with a six point teardown and did not soften it. His first objection is definitional. Compaction cleans up history to keep the agent focused and should be used sparingly when context gets too long, not constantly to keep context small. A filter is not the same job.
The parts that cost you money
The second problem is that the scorer is blind. Theo says the implementation decides per tool call, so a 32k token context model knows very little about what happened before and, in that implementation, does not even see the result of the tool call it is judging. Delete the wrong lines and the agent forgets what it already tried.
Third, the reasoning goes. 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 often drops. Theo says Anthropic requires you to preserve the entire history to get any reasoning data back, so running this inside Claude Code guarantees a dumber model.
Then the bill. Theo says cache writes are the biggest cost by far for agents and that he often sees them go over 60% of his total LLM spend in his own use of Claude Code and Codex. Editing history invalidates the cache for everything after the edit. His example is a history of 1,2,3,4,5,6 where deleting 2 forces a rewrite of 3,4,5,6, which costs more than leaving 2 alone.
He also notes that frontier labs have spent the last year training compaction and long runs into the models themselves, and that in Codex compaction runs on whichever model was previously used in the thread. His advice to anyone tempted is to use the defaults in Claude Code and Codex.
The read underneath is that cheap per call classification is genuinely useful for sorting things that have no shared state, which is where the earlier Jev results came from. An agent thread is the opposite of that. Theo still calls the experiment genuinely interesting, and says the better question is what if harnesses and LLMs never had to think about KV caching at all.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
Deleting these things randomly will keep the model from knowing what it's tried and dooms you to end up in "stupid loops" where the model keeps trying the same thing over and over.
in 2026, why is compaction still a summarization prompt?
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.