APIからエージェントへ：Google Next '26での本当の変化

This is a submission for the Google Cloud NEXT Writing Challenge I watched Google Cloud Next ‘26 thinking I’ll just see better models, faster APIs, maybe some cool demos. But I didn’t expect this. This felt different. Not like “AI is improving”. More like the way we build software is changing. The Moment That Stuck With Me It started simple. JayTee Hazard was creating music. Tina Tarighian was generating visuals. But the interesting part wasn’t the demo. It was what was happening behind it. Gemini was: listening to the music generating code updating visuals in real time And it just kept going. No “run again”. No “generate once” It was a loop. That’s when it clicked for me. This is not prompt to output anymore. This is: input → reasoning → tool use → execution → feedback → repeat The difference is not just output. It’s the system behind it. That loop is the foundation of agentic systems . Then This Number Hit Me Sundar Pichai mentioned: ~75% of new code at Google is AI generated and reviewed by engineers I had to pause there. Not because it’s surprising. But because it confirms something we already feel. We’re not writing everything anymore. We’re: guiding reviewing correcting Almost like we moved from writing functions to reviewing systems. The Part That Felt Real The most interesting part wasn’t the models. It was how they’re actually using this internally. They gave an example of a complex code migration. Instead of one system, they had: a Planning Agent an Orchestrator Agent a Coding agent and Engineers Working together. And they completed it 6x faster . That’s not “AI helping”. That’s a team . So What Does This Mean for Us? This is where things started making sense for me. 1. We’re Not Writing Prompts. We’re Designing Systems With the Agent Development Kit (ADK) , you don’t just create one agent. You define: roles capabilities tool access execution flow Each agent becomes: a stateful unit with memory + tools It felt like building microservices, but instead of APIs you’re wiring intelligence. 2. The API Layer Is Getting Abstracted (MCP) This was subtle but huge. With Model Context Protocol (MCP) inbuilt now: tools expose capabilities in a standard format models understand how to use them context is passed in a structured way Instead of: writing REST calls parsing responses handling retries Your agent does tool invocation via context. Think of MCP as a contract between models and tools 3. Agents Talking to Agents (A2A) With A2A (Agent-to-Agent) . Agents can: discover other agents request capabilities validate outputs Each agent exposes something like: { "name" : "evaluator" , "capabilities" : [ "validate" , "score" , "simulate" ] } And another agent can: evaluator . evaluate ( plan ) This creates dynamic multi-agent coordination. 4. The UI Part Was Unexpected This one felt weird at first. Instead of building dashboards manually. Agents generate UI based on context. Using A2UI : data → structured output output → UI components So instead of build dashboard → connect data. It becomes generate data → UI gets created This flips the flow completely. 5. Memory Makes Agents Actually Useful One of the biggest limitations I’ve felt AI forgets everything With: session state memory bank Agents can: store context recall past decisions refine outputs So instead of stateless prompt → response. You get stateful system → evolving behavior That’s a big shift. 6. DevOps Is Turning Into System-Level Reasoning This part felt unreal. Using Cloud Assist : infra migration → prompt debugging → automated reasoning fixes → suggested patches Under the hood: model + logs + context + tool execution So instead of: checking logs manually tracing errors The system does root-cause reasoning + suggestion What This Means in a Real Project If I think about building something today: Before: write backend connect APIs manage state build UI Now: define agents (planner, executor, validator) connect via A2A use MCP-enabled tools let UI emerge via A2UI The shift is not just speed. It’s how I think about building systems. The Real Takeaway I’m not thinking “AI will replace developers”. I’m thinking the role is changing Before: “How do I write this?” Now: “How do I design a system that can solve this?” And honestly, I’m still figuring out what that means for me. If you’re building with AI right now, are you still writing prompts? Or are you starting to design systems? This is a submission for the Google Cloud NEXT Writing Challenge I watched Google Cloud Next ‘26 thinking I’ll just see better models, faster APIs, maybe some cool demos. But I didn’t expect this. This felt different. Not like “AI is improving”. More like the way we build software is changing. The Moment That Stuck With Me It started simple. JayTee Hazard was creating music. Tina Tarighian was generating visuals. But the interesting part wasn’t the demo. It was what was happening behind it. Gemini was: listening to the music generating code updating visuals in real time And it just kept going. No “run again”. No “generate once” It was a loop. That’s when it clicked for me. This is not prompt to output anymore. This is: input → reasoning → tool use → execution → feedback → repeat The difference is not just output. It’s the system behind it. That loop is the foundation of agentic systems . Then This Number Hit Me Sundar Pichai mentioned: ~75% of new code at Google is AI generated and reviewed by engineers I had to pause there. Not because it’s surprising. But because it confirms something we already feel. We’re not writing everything anymore. We’re: guiding reviewing correcting Almost like we moved from writing functions to reviewing systems. The Part That Felt Real The most interesting part wasn’t the models. It was how they’re actually using this internally. They gave an example of a complex code migration. Instead of one system, they had: a Planning Agent an Orchestrator Agent a Coding agent and Engineers Working together. And they completed it 6x faster . That’s not “AI helping”. That’s a team . So What Does This Mean for Us? This is where things started making sense for me. 1. We’re Not Writing Prompts. We’re Designing Systems With the Agent Development Kit (ADK) , you don’t just create one agent. You define: roles capabilities tool access execution flow Each agent becomes: a stateful unit with memory + tools It felt like building microservices, but instead of APIs you’re wiring intelligence. 2. The API Layer Is Getting Abstracted (MCP) This was subtle but huge. With Model Context Protocol (MCP) inbuilt now: tools expose capabilities in a standard format models understand how to use them context is passed in a structured way Instead of: writing REST calls parsing responses handling retries Your agent does tool invocation via context. Think of MCP as a contract between models and tools 3. Agents Talking to Agents (A2A) With A2A (Agent-to-Agent) . Agents can: discover other agents request capabilities validate outputs Each agent exposes something like: { "name" : "evaluator" , "capabilities" : [ "validate" , "score" , "simulate" ] } And another agent can: evaluator . evaluate ( plan ) This creates dynamic multi-agent coordination. 4. The UI Part Was Unexpected This one felt weird at first. Instead of building dashboards manually. Agents generate UI based on context. Using A2UI : data → structured output output → UI components So instead of build dashboard → connect data. It becomes generate data → UI gets created This flips the flow completely. 5. Memory Makes Agents Actually Useful One of the biggest limitations I’ve felt AI forgets everything With: session state memory bank Agents can: store context recall past decisions refine outputs So instead of stateless prompt → response. You get stateful system → evolving behavior That’s a big shift. 6. DevOps Is Turning Into System-Level Reasoning This part felt unreal. Using Cloud Assist : infra migration → prompt debugging → automated reasoning fixes → suggested patches Under the hood: model + logs + context + tool execution So instead of: checking logs manually tracing errors The system does root-cause reasoning + suggestion What This Means in a Real Project If I think about building something today: Before: write backend connect APIs manage state build UI Now: define agents (planner, executor, validator) connect via A2A use MCP-enabled tools let UI emerge via A2UI The shift is not just speed. It’s how I think about building systems. The Real Takeaway I’m not thinking “AI will replace developers”. I’m thinking the role is changing Before: “How do I write this?” Now: “How do I design a system that can solve this?” And honestly, I’m still figuring out what that means for me. If you’re building with AI right now, are you still writing prompts? Or are you starting to design systems?