Descriptions:
Kath Korevec, product lead on Google Labs’ ADA team, presents Jules — an experimental proactive coding agent built to act on developers’ codebases without waiting to be prompted. The talk reframes the core problem with today’s async agents: even when they handle the execution, developers still carry the mental load of monitoring, following up, and redirecting. Jules is Google Labs’ attempt to eliminate that overhead entirely.
Korevec outlines four requirements for proactive systems: observation (continuous awareness of code changes and workflow patterns), personalization (learning developer preferences including what code is off-limits), timeliness (intervening at natural pauses, not mid-flow), and workflow integration (operating inside terminals, repositories, and IDEs rather than a separate application). She describes a three-level maturity model for proactive agents: level one fixes isolated issues opportunistically; level two learns developer style and anticipates needs; level three — the team’s active focus — understands consequence, connecting code changes to user-facing outcomes using live telemetry and analytics.
At level three, Jules works alongside two other agents: Stitch, a design agent, and Insights, a data and analytics agent. Together, the three propose cross-boundary improvements — such as performance fixes that address UX regressions — driven by real signals rather than static heuristics. Features shipping around the time of the talk include Jules memory, which allows the agent to write and update its own project-specific memories that developers can view and edit. The talk offers a substantive early look at how Google Labs is rethinking developer tooling around always-on, ambient intelligence.
📺 Source: AI Engineer · Published December 13, 2025
🏷️ Format: Keynote Launch
