Next '26: Build Agents with ADK, Skills, and Gemini

Agent Development Kit (ADK) Enables Flexible, Production-Ready Agents

ADK, Google's open-source framework launched at Next '26, stands out for building enterprise agents in 2026. It supports Python (primary), Go, TypeScript, and Java libraries, decoupling agent logic from specific models. Use Gemini 3/3.1 Flash/Pro for reasoning, or integrate Claude, open models on GKE, or any provider. Agents gain intelligence via tools (functions for computation or external services like MCP servers/databases) and skills (new concept: YAML metadata for quick loading + on-demand markdown body with code/scripts).

Skills keep context lean: Agent loads YAML summaries of all skills at startup (e.g., "GIS tool generates marathon routes"), then fetches full body only when needed. This avoids token bloat for complex tasks. ADK 2.0 adds graph-based features for larger agent graphs. Deploy to Agent Runtime, Cloud Run, or GKE for scale.

"When you go to build agent in 2026, you have a lot of options. And we believe that ADK, agent development kit is the best way to do this."

Marathon Planning Demo: Multi-Agent Orchestration in Action

Core demo simulates planning a 10,000-runner Las Vegas marathon via a planner agent in a 3D Las Vegas app (Race Condition repo). Prompt: "Plan a marathon in Las Vegas for 10,000 runners." Agent dynamically loads skills:

GIS Spatial Engineering: Python script processes GeoJSON (Las Vegas road network) to compute exact 42.195km route. Handles constraints: no back-half elevation gains, geofenced to city bounds, water stations at intervals. Math ensures precision—model doesn't hallucinate routes.
```
# Excerpt from skill script
def generate_marathon_route(geojson_data, target_length_km=42.195):
    # Mathematical ops on coordinates for route optimization
    ...
```
Mapping: Queries Google Maps MCP server (natural language over APIs) for places (landmarks like Bellagio, Sphere), weather history (avoid extreme temps).
Race Director: Text-based guidelines from Google Doc (converted via Workspace MCP + Gemini summarization). Covers soft reqs: 3-4 start lanes, porta-potty spacing, traffic impact, economic notes.

Agent iterates: Loads skills on-demand, calls tools, outputs grounded plan. Full code in open-source Race Condition repo (includes .mmd files for Claude/Gemini CLI/Antigravity coding harnesses). Codelab guides setup/deploy.

"We took the task of okay can we take that process marathon planning and make it so that bunch of agents working together can do the same thing if possible even better."

Trade-offs: Skills shine for modular, discoverable capabilities but require upfront YAML curation. Tools handle real-time actions; combine for hybrid intelligence.

Multi-Agent Architectures and Protocols

Post-keynote chats (Ivan Nardini, Casey West) detail Demo 2: Multi-agent setup with real-time evaluation, Agent2Agent (A2A) protocol, A2UI registry. Started Feb '26; evolved from tools to skills differentiation. Identities for marathon: planner + specialized roles (e.g., route optimizer, logistics).

A2A enables agent handoffs; registry discovers skills/UI agents. Built with Vertex AI, Gemini Enterprise Agent Platform. Other segments touch Flutter agents, Firebase SQL Connect (gcloud sql connect), OpenTelemetry tracing, Data Agent Kit, Gemini Nano, Vertex AI Memory Bank.

"We start using tools and then uh we switch and we decide to differentiate between tools and skills."

Developer Resources and Ecosystem

Next '26 emphasizes hands-on: Clone Race Condition for simulation/UI/agents. Use Google Antigravity, Firebase agent skills, Google AI Studio. Hackathons like Gemini Live Agent Challenge; codelabs (e.g., Building Trustable AI at 100 MPH). GEAR hub, 100+ session VODs.

Integrates Workspace MCP (Docs to skills), Maps MCP (NL queries). For trust/scaling: Evaluation loops, memory banks. Opinion: 2026 agents succeed via right tools/skills/runtime—not just models.

"It's not about just okay, what model I choose and what agent framework I use. It's more about how do I give the agent the right tools, the right skills and the right place to run."

Key Takeaways

Start with ADK for multi-language, model-agnostic agents; pair with Gemini for reasoning.
Design skills as YAML metadata + lazy-loaded markdown/code to manage context efficiently.
Ground agents: Use Python scripts for math (GIS routes), MCP for APIs (Maps weather/places).
Clone Race Condition repo; follow codelab to build/deploy marathon planner.
Differentiate tools (actions) vs. skills (discoverable modules); use A2A for orchestration.
Convert docs to skills via Gemini + Workspace MCP for non-deterministic guidelines.
Deploy to Agent Runtime/Cloud Run; trace with OpenTelemetry.
Evaluate Antigravity/Cursor for AI-assisted coding in agent repos.
Join Gemini Live Agent Challenge for hands-on multi-agent practice.