Scoring context with small models: a cheaper alternative to one-shot summarization

marlene_zw · x · 2026-09-23

Instead of one-shot summarizing agent context with Opus or GPT, the author scores each chunk on a 4-level scale across P(still needed for the goal), P(superseded by later activity), and P(needed again) to decide what to keep or discard. Still evaluating output quality, but it's cheaper and faster than one-shot summaries; a video may follow if it pans out.

Related event: Small-Model Probability Scoring for Cheaper RAG and Agent Context Management(3 posts)→

Original post →

More from coding & agent

coding & agent channel →