Gemini Spark's real-world performance falls short of Google's polished demo

Gemini Spark’s real-world performance falls short of Google’s polished demo

Google’s new background AI agent, Gemini Spark, can execute narrowly scoped tasks with surprising competence but falters when asked to handle abstract, multi-step workflows outside the company’s carefully scripted demo scenarios. According to The Verge’s hands-on testing, the gap between what Google demonstrated at I/O and what the agent delivers in a home office setting reveals the persistent brittleness of agentic AI systems in unpredictable real-world conditions.

Email drafting succeeds, but only with retrievable data

When The Verge asked Spark to draft an email compiling monthly grocery spending from a 2026 budget spreadsheet, the agent delivered impressively. Spark located the user’s wife’s email address without explicit instruction, identified the correct budget file in Google Drive despite its unintuitive filename, extracted and averaged monthly grocery totals including incomplete May data, and composed a personalized email with an intimate sign-off phrase. This narrow task — retrieve structured data from known sources and format it into a message — played directly to Spark’s strengths in data integration across Google’s ecosystem.

The baseline for comparison comes from Google VP Josh Woodward’s I/O keynote, where Spark compiled information about Gemini Live launches into a draft email mimicking the user’s tone. That task succeeded because it operated within Google’s own infrastructure, where data provenance and permissions are controlled.

Planning and abstract reasoning remain weak points

Spark’s performance degraded sharply when asked to handle the block party planning scenario Woodward had showcased. The agent created a placeholder friends-and-family table, drafted an email referencing a sign-up sheet that did not exist, and generated presentation slides with formatting issues. When instructed to create the missing sign-up sheet and update the email link, Spark required several minutes of iteration before succeeding — suggesting the agent lacked a coherent plan for the multi-step sequence.

According to The Verge, this failure pattern points to a fundamental limitation: Spark excels at retrieval and data synthesis but struggles with task decomposition, constraint reasoning, and resource creation outside its training distribution. The agent cannot reliably predict whether intermediate steps (like generating a sign-up sheet) are prerequisites for downstream tasks (updating the email).

Cost and privacy tradeoffs raise adoption questions

The Verge frames a deeper skepticism: even when Spark functions correctly, the financial cost and privacy implications of a persistent, background-running agent may outweigh the productivity gains for most users. The agent’s access to email, Drive, contacts, and calendar — necessary for autonomous task execution — creates a large attack surface and data-exposure risk that users must consciously accept.

The phrase “always under your direction” and “designed to check with you before taking major actions” represents Google’s attempt to address mounting concerns about autonomous AI systems. Yet The Verge’s testing confirms that Spark still makes silent mistakes (fabricating resources, leaving tasks incomplete) without always prompting for confirmation, undercutting those safety claims.

Why This Matters

The Gemini Spark review surfaces a critical gap in the agentic AI industry: vendor keynotes showcase cherry-picked workflows that align with model capabilities, while real-world usage patterns expose failures in task planning, constraint handling, and error recovery. For enterprise and consumer adoption of background agents, this reliability gap is disqualifying. Teams evaluating autonomous AI systems must now distinguish between demo-stage performance and production-grade robustness — a distinction that favors narrower, task-specific agents over broad multi-step planners. Google’s willingness to ship Spark despite these limitations suggests the company is willing to trade user trust for early-market positioning, a bet that depends on rapid iteration and visible improvements in task success rates within the next two quarters.