Step 2 · Give the agent its tools (12 min)¶

Goal: Build the log_correlator as a complete ADK agent: a model, an instruction and tools.

Builds on: ADK basics: 1. Agent, 2. Tool, 3. Model, 6. Session state.

The idea: an agent is a model, an instruction and tools¶

AdkAgent(name=..., model=..., instruction=..., tools=[...])

model: MODEL is Gemini when GEMINI_API_KEY is set, and the scripted SimulatedLlm if not.
instruction: the job of the agent. LOG_CORRELATOR_INSTRUCTION is ready in the file.
tools: plain Python functions. ADK reads the type hints and the docstring of each function, and tells the model how to call it. When the model asks for a tool, ADK runs the function and gives the result back to the model.

Look at one tool, query_metrics in sre_agent/src/sre_agent/gcp_tools.py. The docstring is the documentation for the model:

async def query_metrics(filter_expression: str, duration_minutes: int = 15, project_id: str | None = None) -> str:
    """Queries GCP Cloud Monitoring for time series metrics. ..."""

Each tool has an if IS_MOCK: branch. On your laptop, it reads mock_telemetry_data/, not the Google Cloud APIs.

At this time, the log_correlator has a weak instruction and no tools. It cannot check metrics.

Your task¶

Find the TODO:
```
git grep -n "TODO(step-2)"
```
Build log_correlator with AdkAgent(...):
- name="log_correlator" and model=MODEL
- instruction=LOG_CORRELATOR_INSTRUCTION
- tools=[...] with the four tools that the file imports: query_metrics, list_metric_descriptors, analyze_trace_cascade and generate_post_mortem

Run it¶

Do these commands:

uv run workshop/check.py 2
uv run simulate_incident.py --engine-only

Find the tool calls in the log, before the report:

[GCP Observability] Querying mock metrics for filter 'metric.type="run.googleapis.com/container/cpu/utilizations" ...

Look at Observability Metrics in the report. It now shows real values: CPU at 0.24 and 100 database connections. The connection pool is full: this is the cause of the timeout.

With a GEMINI_API_KEY, Gemini decides which tools to call, and writes its own analysis. It usually lists the metrics first (Listing mock metric descriptors in the log), then queries the database connection count. Observability Metrics then shows the connections rising to 100. The instruction tells it to check the metrics: without that sentence, Gemini often calls no tool at all.

Discuss¶

Why does trace_analyzer have no tools? Each tool is one more thing that the model can use incorrectly, and each tool adds tokens to each call. Give each agent the smallest set of tools that it needs. Step 4 applies the same rule to the Orchestrator.
What makes a good tool docstring? Say what the tool returns, when to use it, and give an example of each argument.

Stretch¶

With a key: in LOG_CORRELATOR_INSTRUCTION, remove the sentence that starts with "Always check the metrics". Run again. Does Gemini still call a tool? What does it write about the cause now?
Write a new tool: an async function with type hints and a docstring. Add it to tools.
Share a value through the session state. An agent with output_key writes its answer into the session state (ADK basics, 6):
1. Add output_key="analysis" to log_correlator. Run uv run workshop/basics/try_adk.py. The last line now shows the analysis in the session state.
2. Add output_key="trace_id" to trace_analyzer.
3. In fetch_telemetry, log ctx.state.get("trace_id") next to node_input. Run uv run simulate_incident.py --engine-only, and find the two values in the log. They are the same trace ID: one came through the chain, one through the state.

Stuck? Run uv run workshop/step.py hint 2, or ask the workshop coach in Antigravity. To apply the solution: uv run workshop/step.py solve 2.

Hint and solution¶

Try the step first. Open the hint when you are stuck, and the solution only after the hint.

Hint: where to change the code

File: sre_agent/src/sre_agent/sre_workflow.py, at the module level. The TODO says:

# TODO(step-2): Build the LogCorrelator as an ADK agent: AdkAgent(name=..., model=..., ...).
#   - name: "log_correlator". The workflow and the logs use this name.
#   - model: MODEL. It is Gemini when GEMINI_API_KEY is set, and a scripted model if not.
#   - instruction: LOG_CORRELATOR_INSTRUCTION (above). It tells the model its job.
#   - tools: the four functions imported at the top of this file: query_metrics,
#     list_metric_descriptors, analyze_trace_cascade, generate_post_mortem. ADK reads their
#     type hints and docstrings, and tells the model how to call them.

Solution

--- a/sre_agent/src/sre_agent/sre_workflow.py
+++ b/sre_agent/src/sre_agent/sre_workflow.py
@@ -70,14 +70,12 @@

 trace_analyzer = AdkAgent(name="trace_analyzer", model=MODEL, instruction=TRACE_ANALYZER_INSTRUCTION)

-# TODO(step-2): Build the LogCorrelator as an ADK agent: AdkAgent(name=..., model=..., ...).
-#   - name: "log_correlator". The workflow and the logs use this name.
-#   - model: MODEL. It is Gemini when GEMINI_API_KEY is set, and a scripted model if not.
-#   - instruction: LOG_CORRELATOR_INSTRUCTION (above). It tells the model its job.
-#   - tools: the four functions imported at the top of this file: query_metrics,
-#     list_metric_descriptors, analyze_trace_cascade, generate_post_mortem. ADK reads their
-#     type hints and docstrings, and tells the model how to call them.
-log_correlator = AdkAgent(name="log_correlator", model=MODEL, instruction="Describe the trace.")
+log_correlator = AdkAgent(
+    name="log_correlator",
+    model=MODEL,
+    instruction=LOG_CORRELATOR_INSTRUCTION,
+    tools=[query_metrics, list_metric_descriptors, analyze_trace_cascade, generate_post_mortem],
+)


 @dataclass(frozen=True)

To apply it to your code: uv run workshop/step.py solve 2.