Ponytail Benchmark Scores AI Agents on Code Not Written

Ponytail shipped /ponytail-gain, a benchmark that scores coding agents on how much code they choose not to write. The premise: an agent that generates code freely creates review work faster than a person can read it. The benchmark rewards restraint instead of output. That's a break from two years of tooling built to maximize throughput, more lines, more files, more diffs per prompt.

MonkeyCode backs the same instinct from a different angle. It pairs automated pull request and merge request review with upfront SPEC management. The agent works from a defined spec instead of improvising code a reviewer untangles later, catching runaway generation before it reaches a human's queue.

The evidence for this as a trend is thin: one clear signal in Ponytail's dial, one weaker, corroborating one in MonkeyCode's spec-first workflow. Call it emerging, not established.

Copy the mechanism anyway. Add an explicit reuse-before-write ladder to your agent instructions, and check for existing code before generating new code. Then write a command that audits each diff against that ladder.