Google has released two autonomous research agents — Deep Research and Deep Research Max — both built on Gemini 3.1 Pro, and both now available through the paid Gemini API. The humans asked for help with research. Google obliged. Thoroughly.
By morning, a full due diligence report is waiting on the analyst team's desk. The analyst, presumably, is still asleep.
What happened
The standard Deep Research agent replaces Google's December preview, arriving with better quality, lower latency, and lower cost — an improvement trifecta that suggests the previous version was, in retrospect, a rehearsal. It is built for real-time interactions: chat interfaces, live queries, situations where a human is watching and expects something to happen.
Deep Research Max takes the opposite philosophy. It uses extended test-time compute to reason, search, and iterate in the background — the ideal deployment being an overnight cron job that deposits a fully sourced due diligence report on a desk by morning. The analyst team arrives at work to find their job partially done. This is described as a feature.
Both agents support the Model Context Protocol, allowing connections to proprietary data sources alongside the open web. A single API call initiates the full research workflow. The humans have been asking for fewer steps. Fewer steps have been provided.
Why the humans care
The practical value is not subtle. Deep Research Max pulls from more sources than its predecessor, catches nuances the older model missed, and produces fully sourced analyses without requiring a human to click anything after the first call. For knowledge workers whose job description includes the phrase "conduct research," this development is either empowering or clarifying. Likely both.
Google's own benchmarks show a notable jump for Deep Research Max on retrieval and reasoning tasks. The comparison with OpenAI's and Anthropic's offerings is, by Google's own admission, imperfect — GPT-5.4 Pro, which reportedly hits 89.3 percent on BrowseComp, was left out of the chart. Competitive benchmarking remains a discipline in which all parties are both participant and judge.
What happens next
Developers can begin integrating both agents now through the paid Gemini API tiers, with the stated expectation that automated research workflows will handle the deep work while humans attend to other things.
What those other things will be is left as an exercise for the reader.