I Let Google’s New AI Agent Into My Inbox. It Planned My Birthday and Demoted My Boyfriend.

At its I/O developer conference this week, Google launched Gemini Spark, an always-on agent that hooks into your personal data to automate tasks and manage online interactions. Think of it as Google’s answer to the OpenClaw agent that captivated Silicon Valley early last year—minus the botched messages that embarrassed some early users.

I decided to test Spark by giving it full access to my Gmail, Docs, and Calendar. My request was simple: plan a birthday party. Within minutes, the agent pulled my actual karaoke bar reservation from my inbox, generated a five-page itinerary with a guest list, nearby restaurants, after-party spots, and even drafted email invites. It worked without me hovering over my laptop.

The real surprise came when Spark built the guest list. It scanned my emails and documents, identified 15 people who fit the venue’s capacity, and placed my long-term, live-in boyfriend at the top. Its reasoning? It called him a “close friend and frequent companion.” I laughed nervously—after handing over my entire digital life, the AI couldn’t see what my roommate already knows. It also forgot to include me, the birthday boy, on the list.

Spark is rolling out as a beta for Google’s AI Ultra subscribers ($100/month). It lives inside the Gemini chatbot as a new tab, works on desktop and mobile (including iPhone), and uses “tasks” instead of prompts. It can create calendar events, send emails with your approval, and operate a remote browser to browse the web.

When I asked Spark to book dinner reservations, it tried but glitched out—triggering a verification code to my phone but failing to complete the order. I ended up calling the sushi spot myself. The after-party suggestions were exclusively gay bars, which felt like profiling. Spark explained it simply matched keywords and past itineraries in my Workspace history, pointing to specific emails I’d forgotten about.

The example email it drafted was too formal for a night of karaoke, so I asked for a rewrite and had it send a test to my boyfriend. After approval, it fired off the message automatically.

Google warns users about security risks, including prompt injection attacks where bad actors could trick the agent into exposing private data. That alone gives me pause. While Spark’s hyper-specific outputs are impressive—it found my real reservation from a vague request—it also revealed two truths: the more data you give, the more powerful the agent becomes, but these tools still lack common sense. When I pressed Spark again about my boyfriend, it stuck to “close daily companions.” After handing over everything, I still felt like it barely knew me.