Hands-on
Labs
Build it to understand it. Each lab is a self-contained, runnable walkthrough — step-by-step instructions, real code, and a starter repo you can download. Standalone labs stand on their own; series are meant to be worked through in order.
Hands-on Labs
6 labsAgent-to-Agent Collaboration (A2A)
Wire up the Agent-to-Agent protocol so independent agents can discover each other and collaborate on a task — with a worked, offline Acme example.
Evaluating Agents (Evals as Code)
Treat agent quality like code: build a version-controlled evaluation harness that scores real cases and catches regressions before they ship.
Build an MCP Server
Build a working Model Context Protocol server from scratch and call it from a client — the standard way to give AI agents governed access to your tools and data.
Portable Knowledge for AI with the Open Knowledge Format
Author, validate, consume, and generate institutional knowledge in the Open Knowledge Format — portable AI context that's just markdown files with YAML frontmatter, managed in git.
Discovering and Trusting Agentic Resources with ARD
Stand up both halves of Agentic Resource Discovery offline: publish and cryptographically sign a catalog, run a registry and client, and watch trust verification reject an impersonator.
Building with Google's Interactions API
Build a stateful, agentic workflow on Google's Interactions API — multi-turn server-side state, function calling, tool combinations, and background tasks — in one small Python program.
ADK Series
5 partsYour First Agent with Google's Agent Development Kit
Build your first agent with Google's open-source Agent Development Kit — instructions, a tool, and a local run loop you can actually execute.
Tools, State & Memory
Give your ADK agent real tools, session state, and memory so it can carry context and do useful work across turns.
Multi-Agent Systems
Compose multiple ADK agents into a coordinated system with delegation and sub-agents — the building block for complex workflows.
Evaluation, Callbacks & Safety
Make an ADK agent measurable and production-safe with evaluation, lifecycle callbacks, and guardrails.
Deploy & Connect to the Ecosystem
Deploy an ADK agent and connect it to the wider ecosystem — MCP tools, A2A peers, and a managed runtime.
Gemini Enterprise Agent Platform
11 partsPlatform Tour & Your First Agent in Agent Studio
Get oriented in Google's Gemini Enterprise Agent Platform and build a working agent with no code — instructions, a sample tool, and a live test — in Agent Studio, then clone a ready-made agent from Agent Garden.
Agent Engine: Deploy to the Managed Runtime
Take a code-defined ADK agent and deploy it to Agent Engine, the platform's managed runtime — then call it over the network with managed sessions and watch it in the console. No infrastructure to operate.
Grounding with Vertex AI Search & RAG Engine
Ground the agent in your own documents so its answers are accurate and cite their sources — via a Vertex AI Search data store in the console and a RAG Engine corpus in code, then read the grounding metadata for citations.
Tools & Connectors: APIs, Databases, and MCP
Give the agent hands: connect it to real systems through four tool types — a function tool, an OpenAPI toolset, a governed Integration Connectors data connector, and an external MCP server — all surfaced as ordinary, read-only tools.
Sessions & Memory Bank
Give the agent managed short-term memory (Sessions) and long-term memory (Memory Bank) so it remembers a customer within and across conversations — generating and recalling durable facts without running your own store.
Example Store: Steer Behavior with Few-Shot Examples
Curate few-shot examples centrally in Example Store and retrieve the relevant ones at runtime to steer the agent's behavior — improving quality without retraining a model or bloating the prompt.
Gen AI Evaluation: Measure and Gate Quality
Evaluate the agent with the platform's managed Gen AI Evaluation service — define a dataset, score it with model-based and computation metrics, compare two versions head-to-head, and gate ship/no-ship on the numbers.
Model Armor & Safety Filters
Put a managed safety perimeter in front of the agent: screen user prompts and model responses for prompt injection, jailbreaks, harmful content, malicious URLs, and sensitive data — centrally configured and enforced, not left to the agent's own instructions.
Agent Identity & the Agent Registry
Govern who an agent is and what it may reach: a verifiable SPIFFE/X.509 identity, a namespaced URN in the Agent Registry, an egress policy via Agent Gateway, and a published catalog so agents can discover and trust each other across the org.
Observability, Tracing & Cost
See what deployed agents actually do and what they cost: enable tracing and telemetry, read spans to debug a request, watch the built-in dashboard, alert on errors and latency, and bound spend with budgets and quotas.
Interoperability Capstone: A2A + ADK on the Platform
Tie the managed platform to the open ecosystem: a platform-hosted agent that consumes an MCP server, delegates to a peer over A2A, and is itself exposed over A2A — managed core, open edges, no lock-in.
Microsoft Foundry
12 partsPlatform Tour & Your First Agent
Get oriented in Microsoft Foundry and build a working agent with no code — deploy a model from the catalog, give it instructions and a built-in tool, and test it in the playground.
Foundry Agent Service & the SDK
Build and run the agent from code on the managed Foundry Agent Service — connect with the SDK, create an agent, run it over a thread, and meet the OpenAI-compatible Responses path. Python-primary, with the .NET equivalent noted.
Grounding with Foundry IQ
Ground the agent in your own documents and data so its answers are accurate and cite sources — via File Search (a vector store on Azure AI Search), a connected search index, and Foundry IQ, the unified knowledge layer with agentic retrieval.
Tools: Built-in, Custom, OpenAPI & MCP
Give the agent hands: built-in tools (code interpreter, web search), your own function tools, an entire REST API via OpenAPI, and the 1,400+ tools available through MCP — all surfaced as ordinary, read-only tools.
Threads & Memory
Give the agent durable context: managed threads for a conversation, managed memory across conversations, and bring-your-own thread storage in Azure Cosmos DB so conversation data stays in your own resources.
Evaluations
Measure the agent with Foundry's managed evaluators — define a dataset, run local and hosted evaluation with results linked to the agent trace, compare versions, and gate ship/no-ship on the score.
Content Safety & Prompt Shields
Put a managed safety perimeter around the agent: content filters on the deployment, Prompt Shields for prompt-injection and XPIA, plus PII, groundedness, and protected-material checks — centrally configured, not left to the agent's instructions.
Identity, Access & Governance
Govern who the agent is and what it may reach: a managed Microsoft Entra Agent ID, least-privilege RBAC role assignments, and network isolation — identity and access management, for agents.
Connected Agents & the Microsoft Agent Framework
Compose a multi-agent system: a primary concierge that delegates to focused specialist agents via connected agents — no external orchestrator — built with the Microsoft Agent Framework.
Observability & Cost
See what deployed agents do and what they cost: OpenTelemetry tracing to Application Insights (model calls, tool calls, sub-agent hops), Azure Monitor dashboards and alerts, and budgets and quotas to bound spend.
Fine-tuning & Model Customization
Customize the model itself — fine-tuning, distillation, and GPT-5 RFT — and, just as important, learn when not to. Prepare data, run a job, and prove the customized model beats the base before shipping.
Interoperability Capstone: MCP, A2A & Microsoft 365
Tie a Foundry agent to the open ecosystem and the surfaces your users live in: consume MCP tools, interoperate with peers over A2A, and publish the agent into Microsoft 365 — managed core, open edges.
The Goldwater dispatch
No-hype cloud & AI, in your inbox.
New articles land Monday, Wednesday, and Friday. Get them delivered — no spam, unsubscribe anytime.