Google launches Gemini Spark as a 24/7 autonomous AI agent

Gemini Spark launched today as Google's first explicitly persistent AI agent — running continuously in the background rather than waking up only when a user opens a chat window. Spark watches inboxes, calendars, documents, and (with permission) browser activity, surfacing proactive suggestions and executing multi-step tasks asynchronously. Demis Hassabis framed it at I/O as a 'practice run for AGI' — the first product where the AI's job is to act, not to answer.

Technically, Spark runs on Gemini 3.5 Flash as the default agent model, with escalation to Gemini Omni for multimodal tasks. It uses Google's new Personal Intelligence layer, now expanded to 98 languages, plus a long-running execution sandbox with explicit audit trails for every action it takes against a user's account.

Competitively, this is Google's answer to ChatGPT's agentic mode and Anthropic's Computer Use — but with deeper system integration than either rival can match outside their own apps. Spark can natively call into Gmail, Calendar, Drive, Maps, and Android Auto without OAuth dances. The trade-off is lock-in: switching costs spike once Spark owns your scheduling and inbox triage.

Developer reaction is mixed. Gemini 3.5 Flash impresses on agent benchmarks but devs report it 'collapses into hallucination once grading criteria are added,' and the 5.6x cost premium over base Gemini 3 Flash is hard to justify for non-agentic workloads. The bigger question: DuckDuckGo's 28% traffic surge the same week (696 HN points) suggests a real consumer backlash to AI-injected search — Spark's success depends on whether 'AI does it for you' beats 'AI gets out of my way.'