Day 2 — Kaggle × Google: AI Agents Intensive — Summary 🎬🤖

Hosts

• Smita Kolli — Senior DevRel Engineer, Google Cloud
• Anant Nawalgaria — Co-host
• Guest panel: Alan, Kanchana, Mike, Pierre (Cloud AI) + code labs lead Fran Hinkelman

Week format & logistics 📅

• Deliverables: white papers, companion podcast, hands-on code labs, daily live streams, and AMAs
• Optional capstone project with Kaggle certificates, badges, swag, social recognition
• Course materials delivered by email if registered; also available on Kaggle Learner Portal and announced in Discord
• Free course content; compute/token quotas are limited — upgrading to paid compute avoids quota interruptions (optional)

Day 2 focus — Agents connecting to the world 🌐

• Day 1: What an agent is (model + harness) and shift from syntax to intent
• Day 2: How agents connect to tools, other agents, UIs, and payment systems via open protocols to avoid N×M integration explosion

Key goals:

• Eliminate N-into-M integration technical debt
• Standardize safe, discoverable, interoperable connections across the ecosystem

White paper: “Agent Tools and Interoperability for white coding” — Highlights 📄

• Problem: N models × M tools → O(N×M) integrations → hard to maintain
• Solution: Open protocols that reduce complexity and standardize interfaces

Main protocols explained:

• MCP (Model Context Protocol) — “USB-C” for tool connections
  • Reduces N×M to O(N + M)
  • Supports transports like stdio and Server-Sent Events (SSE)
  • Security-first design (read-only views, role scoping, telemetry)
• A2A (Agent-to-Agent) — lingua franca for agent discovery, coordination, delegation via registries and agent cards
• A2UI (Agent-to-UI) — framework-agnostic spec for agents to declare UI intent using a trusted component catalog (dynamic, safe, component-driven UIs)
• UCP (Universal Commerce Protocol) — merchant/ordering side for autonomous commerce
• AP2 (Agent Payment Protocol) — secure agent payment gateway with human-sign mandates to avoid overspend

Architectural trends:

• Move from single-agent monoliths → distributed multi-agent networks (internal specialization, registries)
• Agents become more like microservices / coordinated subsystems

Security & governance:

• Security-first MCP design (scoped viewer endpoints, read replicas, RBAC, telemetry)
• A2UI allows integration with existing design systems via trusted component catalogs
• Need for FinOps-style token/budget controls and kill switches for runaway loops

Q&A — Key takeaways from panel 🎤

• Google is helping adopt protocols by donating to foundations, providing SDKs/ADKs, CLI tools, and implementing protocols in cloud products (lower barrier to production) — emphasis on standards and scale engineering (stateless transports, large-scale MCP infra). — Mike, others
• UIs will be more personalized and context-aware (A2UI enables dynamic, per-user/per-moment interface rendering while preserving design system governance). — Alan
• DBs should evolve to be agent-friendly:
  • Native MCP exposure (read-only viewers, replicas)
  • Agent-specific RBAC, context-aware payloads, SSE-native endpoints, telemetry and lightweight responses optimized for agents — Kanchana & Pierre
• Career impact for analysts/data roles:
  • Shift from manual data-shoveling to architecting, supervising, and orchestrating agentic pipelines
  • Learn agent tooling, MCP/A2A, and focus on system-level thinking; democratized access to data and tooling. — Mike
• Governance in A2UI:
  • Use trusted component catalogs + existing design systems; A2UI drives components rather than replacing design systems — Alan & Pierre
• Preventing runaway/infinite loops & budget drain:
  • Implement max-iteration caps, anomaly telemetry, FinOps budgets, model-tiering (use cheaper models for routine tasks), caching, and architectural safeguards (kill switches). Prompt efficiency/token optimization matters. — Kanchana & team
• Future breakthroughs:
  • Likely both: improved base models + better agent/harness/tooling and tighter human-in-the-loop feedback loops (loop engineering). Cross-disciplinary model types (beyond autoregressive LLMs) will also matter. — Panel

Code labs (Day 2) — Hands-on 🚀

1. Configure MCP in Antigravity

   • Learn MCP fundamentals and connect Antigravity agent to Google-managed MCP servers (e.g., Developer Knowledge API)
   • Prevent hallucinations by exposing real-time canonical docs via MCP
   • Google provides 50+ managed MCP servers (BigQuery, Maps, Cloud Run, etc.)

2. Antigravity CLI (AGY)

   • Terminal-first agent workflows: planning, tool calling, running agents from CLI
   • Install CLI, run agentic tasks directly from terminal (complements the Agent Manager UI)

Notes:

• Links to labs in emails / Kaggle posts / Discord
• No submission required; experiment end-to-end
• If stuck: prompt Antigravity agent for help or ask Discord moderators

Pop quiz highlights (answers)

• MCP reduces N×M to O(N + M) (linear) ✅
• Transition to internally partitioned sub-agents: Internal specialization ✅
• Protocol for agent negotiation/delegation: A2A ✅
• A2UI definition: Framework-agnostic standard declaring UI intent using trusted components ✅
• Commerce roles: UCP = merchant side; AP2 = payment side ✅

Action items / What to do next ✅

• Run Day 2 code labs end-to-end (MCP + Antigravity CLI)
• Experiment plugging Google-managed MCP servers (BigQuery, Maps, DevKnowledge) into your coding agent
• Join Discord; continue discussion and ask questions (win Kaggle swag for selected questions)
• Prep for Day 3: Agent skills, memory, long-context strategies, and skill specialization

Final quick quotes & vibes ✨

• “Agent = Model + Harness” — harness is critical for real-world usefulness
• “Tokens are the new oil” — manage FinOps, optimize prompts & model usage
• We’re early in the agent era — improvements will come from models, harnesses, tools, and tighter human-agent loops

If you want, I can generate a concise checklist to run the Day 2 code labs (install steps + commands) or extract the exact MCP & AGY CLI commands shown in the video. Which would you prefer?