Compaction without summarization — score the tool calls and drop them

Coding agents still shrink history by asking a language model to summarize. The summary is shorter. It is also a new document: names drift, errors get rounded off, and the original tool output is gone.

Tamara Tran’s post on September 17 treated that as the wrong primitive. If the job is “what still matters for the current task,” you do not need a writer. You need a judge. Her suggested loop scores every tool call and drops the ones that are irrelevant. What stays in context stays verbatim.

That is a System One-shaped question: given the current goal and a blob of Read / Bash / Grep output, is this chunk relevant, and how sure are you? A related project, Winnow, applies the same idea as a calibrated sieve in front of Claude Code.

Jasmine’s quote-tweet of Tran’s post is the broader product question: how many tasks we currently hand to LLMs are actually fast semantic decisions — routing, filtering, scoring, tool selection, escalation? The next generation of agents, in that framing, does not win by putting a chat model in every slot. It wins by picking the cheaper intelligence for the slots that were never generative.

This is still an argument plus early repos, not a published eval that compaction-by-scoring beats compaction-by-summary on long-running coding tasks. The reason it spread on X is that it matches Jev’s constraints instead of fighting them. Jev cannot write a better summary. It can answer “keep or drop?” thousands of times.