AI
Agent-First Debugging via Managed MCP Servers
Denise Kwan used Antigravity’s voice input to ask, “Find out what’s wrong with Dino Quest.” The agent connected to Cloud Run logs through Google-managed MCP servers without API keys or console clicks. It located a 503 error caused by live IO traffic. She then said, “Can you look at the code and fix it?” The agent edited main.py in the leaderboard module. Richard Seroter explained the developer knowledge MCP: it fronts over a dozen Google doc bases and pulls 8-to-12-hour-old snapshots into the coding tool.
Data Agent Kit: Natural Language Database Operations
Christina Lin had built the game at 3:00 a.m. and forgot where she stored Firestore data. She used the Data Agent Kit inside Antigravity to add five treats to all users with a single sentence: “Add five treats for all users.” The agent wrote a Python file, asked for permission, executed it, and updated the Firestore documents. It also added an announcement with a formatted apology message. The skill bundle included the know-how to interact with Firestore without requiring manual document insertion code.
Zero-ETL Analytics with BigQuery Skills
Christina ran a prebuilt skill that automated a zero-ETL analytics pipeline. The skill created a log writer sync to stream live logs into BigQuery. It then used the BigQuery MCP to run parallel queries on the log data. The agent built an interactive dashboard showing traffic overview, request counts, win rate by dinosaur type (Speedy led at 54%), treats per outcome, and top reused dinosaurs. After the dashboard, she asked natural-language questions in Antigravity and drew distribution graphs without SQL or console context switching.
Self-Healing Agents with ADK and Agent-to-Agent Communication
Denise and Christina introduced three Dino agents built with the Agent Development Kit (ADK): a remediation agent, a CI agent, and a CD agent. They hosted all agents on Cloud Run alongside the game for low latency and cost efficiency. The remediation agent listened to Cloud Logging errors via Eventarc and Pub/Sub. It investigated each crash using Gemini, fixed the root cause, opened a GitHub PR, and sent a Slack notification. During the live demo, the audience saw the remediation agent detect a crash and start the fix. The agents communicated using A2A (agent-to-agent) protocol.
Agent CLI and Autonomous CI/CD Pipeline
Christina demonstrated the Agent CLI to scaffold a new T-Rex agent. The command created boilerplate code with agent instructions, tools, test/eval setup, and containerization files. She then triggered the CI agent from Slack with a plain message: “CI agent run CI on branch level two.” The CI agent ran tests, security scans, and built images via Cloud Build and Artifact Registry. It passed results to the CD agent, which assessed a risk score, performed a canary release, monitored for 30 seconds, and deployed to production with auto-rollback on failure. Richard announced upcoming spend caps for Cloud Run and Gemini APIs, now in private preview, that pause traffic when budgets are reached.
Notable Quotes
All of our data and our services from Google Cloud is available for the agent. Denise Kwan · ▶ Watch (4:41)
It’s a mindset shift, right? This isn’t how we’re all used to working. There’s something new here. Richard Seroter · ▶ Watch (11:15)
Nobody wants to get We all know crashes happen at any time, right? Like right now there could be a crash at somebody’s app right now. Denise Kwan · ▶ Watch (16:54)
All you have to do is, oh, I got a Slack notification. You don’t actually have to be at your computer. Denise Kwan · ▶ Watch (19:56)
It’s not sequential, because it thinks for itself. Christina Lin · ▶ Watch (33:11)
Key Takeaways
- Managed MCP servers let agents access Cloud services without API keys or console navigation.
- The Data Agent Kit executes natural language commands against Firestore and other databases.
- Agent-to-agent communication enables autonomous CI/CD with smart decision-making and auto-rollback.