Step 0 · Setup and tour (10 min)

Goal: Install all tools. Look at the system that you complete in this workshop.

1. Check your setup

1. Do these commands:

   git switch -c my-work step-00   # if you haven't already
   uv sync --all-packages
   uv run workshop/check.py 0
   

2. Make sure that each import line shows ok.

GEMINI_API_KEY: not set is not an error.

2. Run the incident

1. Do this command:

   uv run simulate_incident.py
   

2. Read the last lines of the output. You see this line:

   The `diagnose_sre` tool call was blocked by the safety policy (decision: deny).
   

3. Send the request directly to the SRE agent:

   uv run simulate_incident.py --engine-only
   

4. Read the report. It has the cascade table, the Identified Bottleneck (/api/database, 99.3 %) and the post-mortem. Deterministic tools make these parts. The analysis of the ADK agents is not there yet: steps 1 and 2 add it.

The command does these tasks:

• The target app (app/main.py) simulates a Gateway → Backend → Database request.
• The database times out after 10 s.
• The target app writes traces and logs to mock_telemetry_data/.
• The Orchestrator agent gets a request to find the root cause.

The safety policy blocks the tool call on purpose. At this time, the Orchestrator denies all tool calls. You fix this in step 4. Until then, use --engine-only.

With a GEMINI_API_KEY, the Orchestrator is Gemini. It tries the tools, the policy denies each call, and Gemini tells you so in its own words, for example Denied by policy "*".

3. Tour (5 minutes, with the facilitator)

Layer	Location	Function
Target app	app/main.py	Makes the incident: traces, logs and metrics.
Tools	sre_agent/src/sre_agent/gcp_tools.py	Async Python functions that the agents call. The type hints and docstrings become the LLM tool schema. Each tool has an if IS_MOCK: branch that reads local files. analyze_trace_cascade finds the bottleneck span: the span with the most time of its own (exclusive time).
SRE agent	sre_agent/src/sre_agent/sre_workflow.py	ADK multi-agent graph: trace_analyzer → fetch_telemetry → log_correlator (steps 1 and 2). Without an API key, the agents use a scripted model, simulated_llm.py.
SRE agent over A2A	sre_agent/src/sre_agent/a2a_agent.py	The agent card and its skills (step 3). ADK's to_a2a() serves it.
Orchestrator	agent/src/agent/config.py	The Antigravity agent that users talk to. It calls the SRE agent over A2A, with one tool for each skill (step 4). A deny-by-default policy must allow each tool.
A2UI surfaces	sre_agent/src/sre_agent/a2ui_surfaces.py	Makes an A2UI v0.9 surface for each result. A surface contains catalog components, not HTML (step 5).
Chat UI	agent/src/agent/routes.py, index.html, static/sre-a2ui.js	Streams the agent output over SSE. Shows the replies and the A2UI surfaces with @a2ui/lit.

To see your progress at any time, run uv run workshop/step.py status.

✅ Done when: uv run workshop/check.py 0 shows ok for all lines, and you read the two reports.