AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Inside Google Labs: Building the Jules Coding Agent with Jed Borovik

· Latent Space

Recorded at GitHub Universe, this conversation features Jed Borovik from Google Labs discussing Jules, Google's autonomous coding agent. Borovik explains that Google Labs builds innovative AI products in close collaboration with Google DeepMind, with Jules positioned as a coding agent that runs in its own environment—designed for long-running, ambitious tasks rather than local assistance. A central theme is that agent scaffolding has been getting simpler as Gemini models have grown more capable. Borovik says early Jules builds relied on complex sub-agent systems and persona-style multi-agent setups (PM agent, reviewer agent, etc.), but those have largely fallen out of favor. He is similarly skeptical of embedding-based RAG for code, arguing that arbitrary chunk boundaries will never be fully reliable and that grep plus long-context attention tends to win. The discussion also covers context management (Jules supports up to 2M tokens; sessions were extended to 30 days after users kept hitting limits), integrations with Gemini CLI, GitHub Actions, and a Jules API, and the emerging local-to-cloud 'teleport' pattern. Borovik closes by pushing back on the 'vibe coding' label, calling for better specification and verification practices, and expressing bullishness on software engineering jobs via Jevon's Paradox—comparing AI's effect on coding to cheap electricity rather than farm mechanization.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.