ColGrep-Agent-2B cuts Claude Code tokens 34% with a small local search agent

The day's items cluster around one theme: what sits in front of the strong model. ColGrep-Agent-2B gives a concrete number for that idea — a local 2-billion-parameter explorer returned not just candidate files but read files, followed references, and handed line ranges and code back to Claude Code, cutting API tokens 34% and cache-aware costs 40% without a significant accuracy change. The tradeoff is visible too: an 83.3% to 80.8% accuracy dip was not statistically significant, but its confidence interval is wide, so the practical answer offered is a split setup, sending name-and-string lookups to normal ColGrep and behavior questions to the agent version, which held observed accuracy at 83.3% while cutting tokens 21%.

Sakana AI's MASS runs the idea in reverse. Instead of splitting a small model into specialized roles, one shared 27B model is turned into several virtual subagents that solve synthetic research tasks, search for better multi-agent workflows, and evaluate the results; the best team's trajectories are then used to retrain the shared model and re-enter the loop. After two cycles, scores per output token hit 1.2–1.6× the base model across four research evaluations. The paper itself centers the obvious worry: a homogeneous model that shares a blind spot with its evaluator could relearn errors as progress, and it is unclear whether the closed loop stays stable over long runs without external ground truth or a heterogeneous evaluator.

The remaining items supply the plumbing. Claude Code 2.1.296 lets subagents run on a different model than the main conversation and compact earlier, and fixes a bug where a subagent interrupted by session end was mistaken for a user rejection. Python 3.15 adds explicit lazy imports and a sampling profiler, which matter for continuously running agents. In the background, Open Quantum Challenge and Sparseup-fr-v0 both ship an evaluation harness alongside the task or model — the through-line is that once outputs are probabilistic, the scoring loop becomes the thing worth building carefully, and whether the scorer is homogeneous or external looks like the difference that matters in long-term operation.