Deep Research and Deep Research Max: Gemini 3.1 Pro Powers Autonomous Research Agents
Google launched two purpose-built research agents on April 21–22, 2026: Deep Research and Deep Research Max. Both run on Gemini 3.1 Pro, the same model that began rolling out across the Gemini API in March. On the BrowseComp benchmark—a suite of over 1,000 research tasks requiring multi-step web retrieval and reasoning—Gemini 3.1 Pro scores 85.9, approximately 25 points higher than its predecessor Gemini 3 Pro. The agents supersede an earlier research tool released in December 2025.
Deep Research is the efficiency-focused variant, optimised for lower latency and reduced cost compared to the December preview while maintaining higher output quality. It suits interactive use cases where a developer or analyst needs a thorough but timely response. Deep Research Max is the opposite: it allocates more compute per request and targets comprehensive, long-horizon research tasks such as overnight due diligence reports or exhaustive literature reviews triggered by a cron job. This explicit separation of a fast/interactive tier from a thorough/async tier reflects a design pattern emerging across AI products—matching model invocation cost to the urgency and depth requirement of the task.
Both agents support Model Context Protocol (MCP) connections, allowing them to pull from internal company systems, proprietary databases, and specialised financial data sources like FactSet and PitchBook in addition to public web data. Users can upload supplementary files—spreadsheets, PDFs, videos—and review the agent's research plan before execution to guide the direction of the analysis. Both agents generate inline data visualisations as part of their output using HTML or Google's Nano Banana image generator.
The agents are currently available in public preview via paid tiers in the Gemini API through the Interactions API, with enterprise availability coming to Google Cloud. The infrastructure that powers them also underlies research features in the Gemini App, NotebookLM, Google Search, and Google Finance. Developers integrating research automation into their workflows should evaluate the MCP connection story carefully—the ability to point a research agent at internal data sources rather than only public web content significantly expands the surface of useful applications.
Read more — Google Blog
Gemini Enterprise Agent Platform and Agent Builder 2.0
Alongside the hardware and data announcements at Google Cloud Next 2026, Google unveiled the Gemini Enterprise Agent Platform and Agent Builder 2.0. These two products represent Google's enterprise-focused answer to the question of where AI agent lifecycle management should live.
The Gemini Enterprise Agent Platform is positioned as a single control plane for deploying, scaling, governing, and optimising autonomous agents. Rather than building separate orchestration infrastructure for each agentic application, teams using the platform can register agents, define their tool access and data permissions, track invocations and costs, and set policy guardrails in one place. The governance layer is a notable addition: as enterprises move from prototype agents to production agents that make consequential decisions, the ability to audit what an agent did and enforce constraints on what it can do becomes a compliance requirement rather than a nice-to-have.
Agent Builder 2.0 targets the other end of the spectrum: a "low-code, high-logic" environment designed to let non-developers build sophisticated agents using natural language descriptions of desired behaviour. Agent Builder 2.0 generates functional agent configurations from those descriptions, which developers can then review and modify. The pattern is consistent with an emerging division of labour in enterprise AI: product managers and domain experts define the agent's purpose in natural language, while engineers review, harden, and operate the resulting configurations.
Together, these announcements position Google Cloud as a full-stack platform for agentic applications—from model inference through agent orchestration to governance and auditing. The enterprise governance angle distinguishes the Gemini Enterprise Agent Platform from more developer-centric frameworks like LangGraph or CrewAI, where operational concerns are left largely to the application team.
Read more — Oplexa
GGML and llama.cpp Join Hugging Face
In February 2026, Georgi Gerganov and the team behind GGML and llama.cpp joined Hugging Face, marking one of the most significant organisational developments in local AI infrastructure. The announcement frames the move as a combination of two complementary building blocks: llama.cpp for local inference and the Transformers library for model definition, united under a shared institutional home.
llama.cpp has been the dominant open-source inference engine for running quantised large language models on consumer hardware since its debut following the release of the original LLaMA weights. The GGUF format, which llama.cpp helped establish as the standard container for quantised models, is now the preferred default for on-device inference across a wide ecosystem of tools including LM Studio, Ollama, and Jan. By joining Hugging Face, Gerganov's team gains sustainable funding and engineering resources to maintain and grow this infrastructure without depending on community donations or corporate sponsorship that could come with strings attached.
The practical implications for developers are significant. The primary goal is to achieve near "single-click" shipping of new models from the Transformers library directly to llama.cpp-compatible GGUF format. Today, converting a newly released Hugging Face model to GGUF involves a manual conversion script, quantisation choices, and validation—a process that frequently lags behind new model releases by days or weeks. Automated, validated conversion pipelines maintained by the core team would dramatically reduce that gap and ensure quantisation quality is consistent.
The second priority is improving the deployment experience for users who are not comfortable with command-line tools. Better packaging and expanded platform availability—including more straightforward installation paths on Windows and mobile platforms—are on the roadmap. Critically, llama.cpp remains 100% open-source under its existing license with no changes to the project's technical direction or community governance. The team retains full autonomy; Hugging Face is providing resources, not control.
Read more — Hugging Face Blog
Links & Sources
You May Also Enjoy
