Nov 10, 2025 · 43m · latent-space
⚡ Inside Google Labs: Building The Gemini Coding Agent — Jed Borovik, Jules + AIE CODE Preview
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Google Labs engineer Jed Borovik discusses the architectural evolution, capabilities, and product vision of Jules, Google's autonomous coding agent, while exploring how AI is reshaping developer craft, software economics, and engineering communities.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Jed politely resists the host's framing around enforcing matching tests for generated code, asserting that developers should retain full discretion over their repo conventions.
Hardest push from the hosts ▶ 11:56 Declaring embedding chunking fundamentally brokenThe host directly disputes Jed's framing of RAG being just 'hard', asserting that semantic chunking will never be adequate compared to attention and grep.
Biggest teaching moment ▶ 9:41 Explaining the decline of complex agent harnessesJed educates the host on Google Labs' findings that multi-agent personas and complex harnesses act as crutches that become obsolete as foundation models improve.
The host holds their own ▶ 12:05 Demonstrating deep retrieval systems knowledgeThe host articulates the architectural limitations of arbitrary semantic chunking boundaries and cites Cognition's tool-based grep approach.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Jed's AI Journey and Joining Google Labs | 4 | 5 | 1 | 1 | The conversation is genial as Jed explains his transition from Search to Google Labs following the release of Stable Diffusion. The host probes into Google Labs' org chart and internal precursors to Copilot. | |
| Defining Jules: Autonomous and Ambient Coding Agents | 4 | 5 | 1 | 1 | Jed details the product thesis behind Jules as an ambient and autonomous coding agent with its own dedicated compute environment. The host asks clarifying questions about CLI and API integrations. | |
| Simplifying Agent Scaffolds and Moving Beyond RAG | 7 | 4 | 2 | 5 | Jed explains how agent scaffolds have simplified over time as base models improve. The host actively pushes back on embedding-based RAG, arguing chunking is fundamentally flawed and that grep plus attention scales better. | |
| Graduating Jules from Prototype to Production | 7 | 2 | 1 | 1 | The dynamic shifts as the guest asks the host about the upcoming AI Engineer Code Summit. The host takes on an authoritative role, detailing conference dynamics, selective acceptance rates, and hallway tracks. | |
| Long-Context Sessions and the Macroeconomics of Software | 7 | 4 | 2 | 3 | The pair discuss long-context management in 30-day coding sessions and macroeconomic impacts on software engineering. The host invokes Jevons paradox and context compaction techniques, matching Jed's industry outlook. | |
| Beyond Vibe Coding: Interactive Planning and Multimodality | 7 | 4 | 2 | 3 | The discussion covers moving beyond 'vibe coding' towards spec-driven and interactive planning development. The host pitches Cognition's interactive planning philosophy and multimodal video inputs, while Jed discusses test-generation trade-offs. | |
| Conference Connecting and Google Labs Recruitment | 3 | 2 | 0 | 0 | A friendly wrap-up segment where Jed invites conference attendees to share their workflows, criticisms, and interest in joining Google Labs. |