Our Take
Google is shipping agents across consumer products before the infrastructure to run them reliably at scale is proven; execution risk is real.
Why it matters
Agent deployment in high-traffic surfaces (Search, Gmail) exposes the gap between lab demos and production stability. If adoption stalls on latency or error rates, the entire agent-first narrative deflates.
Do this week
Developers: Test Information agents in Search and Gemini Spark against your actual query patterns before committing to agent-first architecture; document failure modes and latency percentiles.
Google announces Gemini Omni, 3.5 Flash, and agent deployments
Google released two new models at I/O 2026: Gemini Omni, capable of multimodal input and output (starting with video), and Gemini 3.5 Flash, positioned as the first model combining frontier intelligence with action. The company also unveiled advancements to Google Antigravity, its agent-first development platform, and rolled out agentic experiences across consumer products.
New agent deployments include Information agents in Search, Gemini Spark and Daily Brief in the Gemini app, and Universal Cart, described as an intelligent shopping cart. Google also announced expansion of Gemini across products including Google Pics, intelligent eyewear, and Ask YouTube.
Agents in Search mean Google is betting on user trust at scale
Search agents are not new. But deploying them as a default experience in Google Search is a structural bet: that users will accept agents making decisions (filtering results, recommending refinements, executing tasks) within the world's highest-traffic information surface. No benchmarks comparing agent error rates to traditional search results are published. No independent testing of latency impact on p95 queries is disclosed.
If Information agents reduce query success rates or increase session friction, the cost is measurable: user session time, advertiser confidence, and competitive pressure from OpenAI's Search, which launched without agent-first positioning. Google is moving faster than it is publishing safety or reliability data.
Audit agent readiness before adoption
If you are building applications on Gemini Omni or 3.5 Flash, request production SLAs for agent performance: error rates per task type, p99 latency, and rollback procedures. Test Gemini Spark agents in your domain before integrating them into customer-facing flows. Ask your Google contact explicitly whether Information agents in Search are A/B tested and what the fallback behavior is when an agent fails. Do not assume that Google's internal deployment success predicts your own.