Cursor tracks agent code keep rate

// 1d agoPRODUCT UPDATE

Cursor tracks agent code keep rate

Tibor Tee, a community developer at Cursor, shared that the team tracks the "Keep Rate," the percentage of agent-generated code remaining in a user's codebase over time. This telemetry helps evaluate the long-term quality and correctness of AI agent output.

// ANALYSIS

Code volume is a vanity metric; "Keep Rate" is the ultimate measure of AI agent utility and codebase hygiene.

–High AI code acceptance rate means nothing if developers must immediately refactor or delete the generated changes.
–Tracking survival rates over fixed time intervals exposes the long-term maintainability of agentic output.
–This telemetry allows the Cursor team to continuously run online A/B testing and calibrate their models based on actual human retention.
–The metric incentivizes building agents that prioritize code correctness over sheer output speed.

// TAGS

cursoragentcode-retentiontelemetrydeveloper-productivitysoftware-engineering

DISCOVERED

1d ago

2026-06-20

PUBLISHED

1d ago

2026-06-20

RELEVANCE

8/ 10

AUTHOR

tibor_tee

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS1h ago

OpenClaw reaches its strongest week of activity after transitioning to a non-profit structure and improving software quality.

Creator Peter Steinberger shared that despite the initial hype dying down, OpenClaw has improved quality, expanded its team, and registered its strongest week of adoption so far. Steinberger highlights the project's transition to a non-profit foundation, contrasting its mission with venture-backed competitors that prioritize commercial interests.

BENCHMARK1h ago

Claude Opus outperforms GLM-5.2 in coding

A head-to-head evaluation prompting GLM-5.2 and Claude Opus to build a 3D WebGL platformer from scratch showed Opus completing the task in half the time with fewer bugs. While GLM-5.2 is a cost-effective open-weights alternative, the test highlighted the advantage of Opus's multimodal capabilities in using screenshots to self-correct visual bugs.

MODEL1h ago

Sakana AI launches Fugu orchestration API

Sakana AI has launched Sakana Fugu and its high-performance variant, Fugu Ultra, transitioning the multi-agent orchestration system from beta to full commercial availability. Operating via a single OpenAI-compatible API, Fugu dynamically coordinates tasks across a pool of diverse frontier models to handle complex reasoning while helping developers avoid single-vendor lock-in.