Issue No. 10 • June 2, 2026
Gradient Brief
MLOps & AI Infrastructure — for the engineers building it
Google I/O 2026 Rewires the AI Developer Stack
Google I/O 2026 kicked off today, and the developer announcements signal a massive shift in how Google expects ML engineers to build and deploy agentic applications. The most significant release is the Gemma 4 model family. Unlike previous generations that carried restrictive licenses, all four Gemma 4 models (E2B, E4B, 26B-A4B, and 31B) are released under the Apache 2.0 license. The 26B Mixture-of-Experts (MoE) variant is particularly notable for local inference: it requires only 3.8 billion active parameters per token while hitting near-frontier benchmarks, and supports a 256K context window. The models feature explicit <|think|> token control for hybrid reasoning and are already available for local deployment via tools like llama.cpp and Unsloth.
On the orchestration side, Google announced the general availability of the Agent Development Kit (ADK) v2.0 and the Agent-to-Agent (A2A) Protocol v1.2. ADK v2.0 introduces a new graph-based Workflow Runtime for composing deterministic execution flows, replacing the looser orchestration of the 1.x series. The A2A protocol update is arguably more important for the broader ecosystem: IBM's competing ACP protocol has officially merged into A2A, and the standard is now running in production across more than 150 organizations, including Microsoft, AWS, and Salesforce. The update also introduces Event Compaction, which reduces token usage by 38% and latency by 18% in multi-agent communication.
Finally, Google is aggressively restructuring its developer tooling around outcome-driven agents. The company announced the sunset of Firebase Studio (formerly Project IDX), with no new workspaces allowed after June 22, 2026. It is being replaced by Google Antigravity, an agent-first desktop IDE. Alongside this, Google detailed Jules V2 (internally "Project Jitro"), an autonomous coding agent running on Gemini 2.5 Pro that targets specific outcomes (e.g., "raise test coverage to 80%") rather than line-by-line instructions. Jules V2 operates within secure Google Cloud VMs and processed over 140,000 code improvements during its beta phase.
Tool of the Week: Confluent MCP Server & Agent Skills
Open Source (Apache 2.0) | Local or Managed | Confluent
An open-source Model Context Protocol (MCP) server and domain knowledge package that gives AI coding assistants direct, governed access to Apache Kafka and Confluent environments.
ML pipelines increasingly depend on Kafka for real-time feature stores, model serving event streams, and training data ingestion. Until now, AI coding assistants have struggled with streaming infrastructure, often writing generic Kafka code that ignores Schema Registry requirements, Flink SQL patterns, and Change Data Capture (CDC) best practices. Confluent's release solves this by providing both the operational access (via the MCP server) and the domain expertise (via Agent Skills).
The local MCP server works with both Confluent Cloud and local Docker Kafka clusters. AI assistants can discover topics, deploy Flink SQL statements, configure connectors, and query metrics directly from the editor. Every tool is annotated with input schemas and operational context, allowing the AI to assess the impact of an action before executing it. A zero-configuration managed version is also available within Confluent Cloud.
Quick Hits
- Dell Technologies World 2026 Expands AI Factory Dell announced massive expansions to its AI Factory, including the new PowerRack (fully integrated compute, networking, and storage) and Dell Deskside Agentic AI powered by NVIDIA NemoClaw. Most notably for enterprise ML teams, Dell is bringing OpenAI Codex, Palantir Foundry, and Google Gemini 3 Flash on-premises via Dell infrastructure.
- Dust Raises $40M Series B for "Multiplayer AI" The enterprise AI platform raised $40 million led by Abstract and Sequoia, with participation from Snowflake and Datadog. Dust is betting heavily on "multiplayer AI"—shared workspaces where humans and agents collaborate in parallel with persistent context, rather than isolated single-player chatbot sessions. The platform is currently deployed across 3,000 organizations.
- Firebase Studio Sunset Warning If your team built workflows around Firebase Studio (the cloud IDE formerly known as Project IDX), the migration clock is ticking. Google announced today that no new workspaces can be created after June 22, 2026, with a full shutdown scheduled for March 22, 2027.
Gradient Brief is published for ML engineers, data scientists, and technical founders. Forward to a colleague who should be reading this.