Please enable JavaScript.
Coggle requires JavaScript to display documents.
:mortar_board: 5-Day AI Agents Intensive by Google & Kaggle - Coggle…
:mortar_board: 5-Day AI Agents Intensive by Google & Kaggle
Day 1: Agentic SDLC
:dizzy: Vibe Coding vs Agentic Engineering
85% devs use AI, 41% of new code is AI-generated
Vibe Coding: trial & error, eye test, BBQ
Agentic Engineering: specs, tests, evals, Michelin kitchen
:busts_in_silhouette: AI Agent Formula
Agent = Model (10%) + Harness (90%)
Harness: Sandboxes, Guardrails, Orchestration, Observability
:package: 6 Types of Context
Instructions
Knowledge
Memory
Examples
Tools
Guardrails
:conductor: Conductor vs Orchestrator
Conductor: real-time pair coding in IDE
Orchestrator: async swarm, PR review
:moneybag: Token Economics
CapEx (setup) vs OpEx (runtime cost)
Vibe: low CapEx, high OpEx
Agentic: high CapEx, low OpEx
Model Routing: cheap model for simple tasks
:chart_with_upwards_trend: The 80% Problem
AI solves first 80% instantly
Last 20% needs human intuition
Day 2: Tools & Interoperability
:link: MCP — Model Context Protocol
Standard protocol for tool integration
Eliminates duplicate integration code
MCP Servers: filesystem, DB, web, custom
:arrows_counterclockwise: A2A Protocol
Agent-to-Agent communication standard
Orchestrating multi-agent swarms
Version 1.0 released
:frame_photo: A2UI — Generative UI
Agent sends declarative JSON
Client renders safe interactive UI
:wrench: Google ADK — Agent Development Kit
Built-in MCP + A2A support
Day 3: Agent Skills
:book: SKILL.md Anatomy
Formal skill passport for agents
Contains: name, description, tools, tests, examples
Dynamic context on-demand
:test_tube: Evals & Testing
Deterministic tests for code
Systematic evals for AI reasoning
:warning: Context Rot Prevention
Static context (heavy backpack): rules, always loaded
Dynamic context (map in pocket): skills, on-demand
Day 4: Agent Quality & Evaluation
:mag: Observability
Tracing: full call chain visualization
Logging: structured logs, stack traces
Metrics: latency, cost, accuracy
:checkered_flag: Quality Gates
Automated evaluation pipelines
Regression detection
Hallucination guards
Day 5: Prototype to Production
:cloud: Deployment
Google Cloud Run for MCP servers
Docker containerization
CI/CD pipelines
:floppy_disk: Memory & Vector DB
Chroma, Pinecone, Vertex AI Vector Search
Long-term agent memory
:trophy: Capstone Project
Combine: Gemini + MCP + Skills + Evals
Full-stack agent system
:link: Key Protocols
MCP: tool integration standard
A2A: agent orchestration
A2UI: generative UI
ADK: Google dev kit
:warning: Best Practices
Always validate with evals
Use sandboxes for safety
Static context in agents.md
Skills for dynamic context
Model routing to optimize cost
:bar_chart: Progress: =(COUNT(LEAVES)/52*100)%
=(1)