The Four-Layer Agent Ladder

Google has introduced a structured, four-layer approach to building production-ready AI agents, designed to allow teams to start with low-code tools and scale to full engineering control without platform lock-in. All layers are connected by the Agent2Agent (A2A) Protocol, ensuring interoperability across the stack.

  1. Agent Studio (Low-Code): A visual, UI-first environment for rapid prototyping, allowing developers to convert visual workflows into code.
  2. Managed Agents API: A fully hosted, configuration-first approach for developers who need persistent agents that can reason, execute code, and call tools without managing infrastructure.
  3. Anti-Gravity 2.0: An agent-first development platform that acts as "mission control" for orchestrating cohorts of agents. It supports dynamic sub-agents and background automation, with a desktop app, CLI, and SDK for flexible deployment.
  4. Agent Development Kit (ADK) 2.0: The engineering-first layer providing full code-level control. It has shifted from an imperative model to a graph-based workflow, allowing developers to define agent steps as a graph. This supports both deterministic and creative tasks, with a slider to balance the two.

Core Infrastructure and Governance

To support these agents, Google has introduced primitives that solve common production challenges:

  • Skills Registry: A private, organization-scoped repository for agent skills (instructions and tools defined in markdown). It features dynamic discovery, where agents load only the necessary skills at runtime, resulting in leaner, more efficient agents.
  • Agent Identity & Governance: Tools to track agent activity and limit the "blast radius" of their actions. This includes the Agent Payment Protocol (AP2), which allows developers to set spending mandates and limits for agents performing commercial tasks.
  • Agent CLI: A unified tool for scaffolding, local testing, evaluation, and deployment to the managed agent platform, eliminating the need to manage disparate commands.

Model and Multimodal Advancements

  • Gemini 3.5 Flash: The new default model, co-designed with Google’s TPUs. It is optimized for long-horizon agentic tasks, offering significant speed improvements and benchmark performance that exceeds previous Pro models.
  • Gemini Omni Flash: A multimodal creation model capable of fusing text, images, audio, and video into coherent video clips. Unlike standard generation models, Omni reasons about the input physics and context, allowing for consistent scene generation. It includes built-in watermarking via SynthID for provenance tracking.