Giving agents “hands” and a common language to plug things in
Day 1 was about what agents are. Day 2 is about how they touch the real world — databases, websites, other software, and even other agents — without writing a separate custom connector every time.
The core lesson: standards beat one-off hacks. When everyone uses the same “plug shape,” tools swap easily, like USB devices.
A defined way for one program to ask another program for data or actions. Example: “send this JSON, get weather back.” Agents use APIs to read calendars, query databases, or post messages.
Any capability you give the agent: run code, search docs, create a ticket, fetch a file. Often exposed as functions the LLM can choose to call.
You type commands in a terminal window. No mouse-heavy UI — good for scripts, servers, and repeating tasks. agy in this course is Antigravity’s CLI.
GUI = steering wheel and dashboard. CLI = speaking short precise orders to the machine.
Different products working together because they follow the same rules. Your agent app should not need a totally new integration for every model and every service.
An agreed public rulebook for how systems talk. Day 2 introduces several agent-related protocols (MCP, A2A, etc.).
A standard way for an AI client to discover and use tools from an MCP server. One MCP server can serve many clients; one client can plug in many MCP servers.
USB-C for AI tools — same connector, many devices.
Agents advertising what they can do and asking other specialized agents to help with subtasks — like delegating within a team.
Agents outputting UI pieces (forms, buttons, layouts) in a safe, structured way instead of random HTML that might be unsafe.
If you have 5 AI apps and 10 tools and wire each pair manually, you maintain 50 integrations. Protocols reduce that mess to “each app speaks MCP” + “each tool offers MCP.”
| Category | What it does | In this course | Same idea elsewhere |
|---|---|---|---|
| AI coding CLI | Run an coding agent from the terminal; scriptable and fast for power users. | Antigravity CLI (agy) |
Claude Code, Gemini CLI, GitHub Copilot CLI, Aider |
| Tool protocol / plug standard | Standard socket between AI apps and tool servers. | MCP | OpenAPI tools, LangChain tools, custom function calling (non-standard) |
| MCP server | Exposes a service (docs, DB, tickets) in MCP format. | Google Developer Knowledge MCP | Filesystem MCP, GitHub MCP, Postgres MCP, company-internal MCP servers |
| Multi-agent coordination | Agents delegate work to other agents. | A2A (concept) | CrewAI crews, AutoGen groups, OpenAI Swarm patterns |
| Generative UI | UI adapted to the user or task. | A2UI (concept) | Custom React from LLM, Vercel v0 UI output (with your own safety rules) |
What you will get from it: Why open protocols matter for vibe coders, and how they let you assemble a “virtual team” of tools without drowning in custom glue code.
Main ideas: MCP for tools, A2A for agent teams, A2UI for interfaces, and commerce protocols (UCP/AP2) for shopping and payments with human approval on money moves.
Concept practiced: Terminal-based agent use — permissions, models, interactive vs one-shot prompts (agy -p "...").
Transferable skill: Every AI CLI has trust folders, model pickers, and permission modes; learn to read those settings anywhere.
Codelab 2: Google Developer Knowledge MCP in Antigravity codelabs.developers.google.com/developer-knowledge-mcp-antigravityConcept practiced: Wiring an MCP server into a client via config file — API key, server URL, refresh, test with prompts.
Transferable skill: Same steps for any MCP host (Cursor, Claude Desktop, VS Code MCP extensions): add server entry → authenticate → verify tools appear.
Experts explain MCP in production, database security for MCP, and personalized UI futures. Includes codelab demos and a pop quiz.