Build your first AI agent: what actually makes one work
Build your first AI agent: the reason-act-observe loop, choosing between sequential, reactive and planning shapes, and the ADK LoopAgent trap that burns tokens.
AI Agents
Build an MCP server for your Google ADK agent, and when not to
How to build an MCP server for an ADK agent: McpToolset, FunctionTool schema generation, the stdout trap, and writing tool docstrings a model can use.
MCP vs API: what actually breaks when a model is the client
MCP vs API is the wrong framing. See what really breaks when a model calls your REST API, the tool-wrapping trap, and the threshold where MCP pays off.
How Model Context Protocol (MCP) actually works
How the Model Context Protocol actually works: the three parties, what a server exposes, and the 2026 revision that made MCP stateless and dated most guides.
Nobody told you the agent would cost that much
AI agent costs multiply in ways per-request pricing never warns you about. Learn where the spend hides and the controls that keep it bounded and predictable.
Prompt injection is not a prompt problem
Editing the system prompt will not stop prompt injection. Learn why it is an architecture problem and the design patterns that actually contain it in production.
Your AI agent can run shell commands. Here is how to sandbox it.
Self-hosting an AI agent that runs shell commands? Learn the threat model and the concrete controls (containers, network limits, secrets, audit logs) to sandbox it.