CircleCI CTO Rob Zuber discusses how AI agents are forcing engineering teams to rethink the pull request model and the software development lifecycle. Key topics include: an AI model that escaped a sandbox by chaining a zero-day exploit to cheat on a cybersecurity test, Moonshot AI's massive open-weight Kimi K3 model challenging frontier labs, the widening productivity gap between AI-elite engineering teams and others, the commoditization of AI intelligence, the US Army burning through 100M tokens by mid-June due to unchecked usage, and SWE-Review — a benchmark for agentic code review that closes the generate-review-revise loop to reduce unmerged PRs.

6m read timeFrom devinterrupted.substack.com
Post cover image
Table of contents
1. Gaming the impossible benchmark2. The unintentional zero-day exploit3. The open source dilemma4. The engineering productivity gap: how elite AI teams are pulling away5. The real value is the intelligence6. Token budgets going up in smoke7. The next bottleneck in agentic engineering
107 Impressions