Google: reusable skills and A2A chatter
Coverage this weekend emphasizes Google’s push toward reusable AI 'skills' and the A2A protocol for agent-to-agent orchestration, describing these as patterns for composing multi-agent workflows and repeatable tasks (mainlandmoment.com) (ouranimeworld.com). The accounts treat skills as modular units that can be invoked across environments like Chrome and cloud services (mainlandmoment.com) (ouranimeworld.com).
Google is pushing two building blocks for artificial intelligence agents: reusable “skills” for repeatable jobs and Agent2Agent, or A2A, for agents that need to talk to each other. (developers.googleblog.com) A skill is a packaged set of instructions and resources that an agent can load when needed instead of carrying every rule in one giant prompt. In an April 1, 2026 post, Google said its Agent Development Kit, or ADK, now includes a SkillToolset with four patterns, including file-based skills and a “skill factory” that lets an agent generate new skills at runtime. (developers.googleblog.com) A2A is the second piece: a common language for separate agents to discover each other, describe what they can do, and exchange messages securely. Google introduced the protocol in April 2025 and said it was designed to complement Anthropic’s Model Context Protocol, which connects an agent to tools and data sources. (developers.googleblog.com) Google’s recent developer guides frame the split this way: use a tool-like agent for an isolated, reusable task, and use a sub-agent or remote agent when the job needs its own context or a longer back-and-forth. A December 2025 Google Cloud post said tool-style agents are best when the calling agent only needs a reliable result and does not need the internal reasoning. (cloud.google.com) That distinction lines up with Google’s broader push away from one giant assistant and toward teams of smaller agents. In a January 2026 guide, Google described eight multi-agent patterns in ADK, including a sequential pipeline, a router, and a human-in-the-loop setup for production systems. (developers.googleblog.com) Google is also wiring those patterns into products beyond developer demos. When it launched Agentspace in December 2024, Google said employees would be able to access enterprise search and agents from Chrome, and later updates added a no-code Agent Designer and Agent Gallery for discovery and reuse. (cloud.google.com 1) (cloud.google.com 2) On the plumbing side, Google has been turning standalone ADK agents into networked components. A July 2025 Google Cloud guide showed how to convert an ADK agent that uses Model Context Protocol tools into an A2A-compatible service that can join a larger multi-agent system. (cloud.google.com) Google and its partners have also been trying to move A2A beyond Google-only software. Google’s documentation says the protocol was donated to the Linux Foundation, and Google’s January 2026 A2UI announcement said agents from Google, Cisco, International Business Machines, SAP, and Salesforce were already using the approach to collaborate across trust boundaries. (a2a-protocol.org) (developers.googleblog.com) The pitch is straightforward: keep expertise in small reusable packages, then let specialized agents call each other when one package is not enough. Google’s own examples now span coding help, enterprise search, browser access, and cloud microservices, all built around the same idea of composing repeatable parts instead of stuffing everything into one model session. (developers.googleblog.com) (blog.google) (cloud.google.com)
