🤖

AI Engineering — SDLC, Agents & MCP

How AI systems actually get built and operated: the AI-native SDLC, agent harness engineering, and the Model Context Protocol that wires agents to tools. Agent = Model + Harness, and the harness is roughly 90% of the system — so most agent failures are configuration failures, not model limitations. Decks for the concepts, browser drills for the real JSON, and labs that need no API key and no real model.

Start here

MCP Fundamentals15 cards — step 1 of 8

Then get hands-on: Agent harness lab — 9 scenarios.

Flashcard decks

8 decks · 152 cards · in study order

Model Context Protocol

6 decks · 89 cards

1MCP FundamentalsWhat MCP is and why it exists: the 'USB-C for AI' problem, the host / client / server roles, the JSON-RPC 2.0 foundation, date-based protocol versioning, and how it differs from bespoke per-app integrations.15 2MCP — Big Picture & EcosystemHow MCP fits the real world: client config files (mcpServers), the host/registry ecosystem, when to build a server vs use function calling or a plugin, a worked case study (an ai-model-selector MCP server), and the throughlines that tie the whole protocol together.14 3Server Primitives — Tools, Resources, PromptsThe three things an MCP server exposes: TOOLS (model-controlled actions), RESOURCES (application-controlled data), and PROMPTS (user-controlled templates) — their JSON shapes, list/call/read methods, and the control model behind each.16 4Client Primitives — Sampling, Roots, Elicitation, TasksWhat the CLIENT offers back to the server: SAMPLING (borrow the host's LLM), ROOTS (expose filesystem boundaries), ELICITATION (ask the user for input, form or url), and TASKS (long-running async operations, added in 2025-11-25) — plus the human-in-the-loop principle.14 5Transports & LifecycleHow messages actually move and how a session begins: the stdio transport (local subprocess) and Streamable HTTP (remote), the deprecated HTTP+SSE, the initialize handshake and version negotiation, the initialized notification, and the utility flows (ping, cancellation, progress).15 6Building Servers & SecurityFrom SDK to safe deployment: the official SDKs and how a server is structured, the OAuth 2.1 authorization model for remote servers (resource server, protected-resource metadata, PKCE, resource indicators), and the threats — tool poisoning, prompt injection, confused deputy — and the trust-and-consent principles that counter them.15

The AI-native SDLC & agent harnesses

2 decks · 63 cards

7The AI-Native SDLCHow software development changed: the four layers from prompt to loop engineering, Agent = Model + Harness, the six context types, Conductor vs Orchestrator, and which SDLC phases actually compressed. Built from Google/Kaggle's June 2026 whitepaper. Verified 2026-07-31.33 8Agent Harness EngineeringThe practical build order for the code that wraps a model: the five extension points, what belongs in an instructions file and what doesn't, context budget, LSP, subagents, and the anti-patterns. Built from Anthropic's large-codebase guide. Verified 2026-07-31.30

Practice exams

1 set

MCP FundamentalsDraws a random 30-question set each attempt from a pool spanning MCP fundamentals, the server primitives (tools/resources/prompts), the client primitives (sampling/roots/elicitation/tasks), transports & lifecycle, and building & security. Scores you, shows what you missed with explanations, and breaks down weak areas by topic. Progress saves in your browser — no login. Authored from the public MCP specification (rev 2025-11-25).

Agent harness lab

live

Build the 90% that isn't the model. A deterministic badly-behaved stub lets you write extraction, validation, bounded retry, tool dispatch and evals — then fix harnesses that swallow errors, loop forever, or ship wrong answers. No API key needed.

MCP drills

11 drills

Write real MCP JSON — mcpServers client config (local stdio and remote HTTP servers), tool/resource/prompt definitions, and JSON-RPC messages (the initialize handshake, capability declarations, tools/call) — graded live in your browser against the actual shapes (spec rev 2025-11-25).

Curated resources

verified July 2026

Official docs lead — the MCP spec moves fast, so most third-party videos go stale quickly. One vetted course (Anthropic × DeepLearning.AI) is included.