Basics: ADK, A2A, A2UI and the Antigravity SDK

The workshop uses four technologies. Each page below teaches the basics in about 5 minutes, with code from this repository and one command to try. The workshop steps build on these basics.

Read the pages before the workshop, or when a step links to them.

Page	What you learn	Try it	Steps
ADK	Build agents: models, instructions, tools and workflows	uv run workshop/basics/try_adk.py	1, 2
A2A	Let agents talk to other agents: agent cards, skills, tasks	uv run workshop/basics/try_a2a.py	3, 4
Antigravity SDK	Run an agent with safe limits: policies and hooks	uv run workshop/basics/try_policy.py	4
A2UI	Let agents send UI: surfaces, components, data and actions	The playground: http://localhost:8080/playground	5

Before you try the commands, run uv run simulate_incident.py --engine-only one time. It makes an incident for the agents to find.

The big picture

flowchart TB
    U["🧑 You, in the chat"]
    O["🤖 Orchestrator<br/>Antigravity SDK<br/>policy: deny *, allow 3 tools"]
    S["🩺 SRE agent · ADK<br/>trace_analyzer<br/>→ fetch_telemetry<br/>→ log_correlator"]
    T[("traces · logs · metrics")]
    U <-->|"A2UI: surfaces and actions"| O
    O <-->|"A2A: agent card, then messages"| S
    S -->|tools| T

• The Antigravity SDK runs the Orchestrator: the agent that you talk to. Its policy lets it call only three tools.
• A2A connects the Orchestrator to the SRE agent. They are separate programs, and can be separate services.
• ADK builds the SRE agent: two model-driven agents and a Python step in a workflow.
• A2UI carries the result back as UI: cards, tabs, badges and buttons, not HTML.

One click, from start to end

You click Diagnose on an incident in the chat. This is what happens:

1. A2UI action. The button sends an event: {"name": "diagnose_incident", "context": {"traceId": "…"}}. The chat changes it into a new turn: "Diagnose trace …".
2. Antigravity. The Orchestrator's model chooses the tool diagnose_sre. The runtime checks the policy: allow("diagnose_sre"), so the call runs.
3. A2A request. The tool reads the SRE agent's card at /.well-known/agent-card.json, and sends a JSON-RPC SendStreamingMessage with the metadata skill: diagnose_incident, the trace ID, and the A2UI client capabilities of the chat.
4. ADK workflow. The SRE agent runs its workflow: the trace analyzer picks the trace, a Python step gets the spans and logs, and the log correlator calls tools and writes the root cause.
5. A2A stream. While the workflow runs, the SRE agent sends status updates ("Fetching traces…"). The chat shows them as progress. At the end, the task has an artifact: the report as text and as A2UI data parts.
6. A2UI render. The chat renders the surface with @a2ui/lit: a card with a severity badge, tabs and buttons. Each button can start the next click.

Where to learn more

Technology	Official documentation
ADK	https://google.github.io/adk-docs/
A2A	https://a2a-protocol.org
A2UI	https://a2ui.org
Antigravity	https://antigravity.google/docs