Skip to content

Step 3 · Publish an A2A skill (12 min)

Goal: Make the SRE agent offer a third skill, write_post_mortem, to other agents.

Builds on: A2A basics: 1. Agent card, 2. Skill, 7. Server and client.

The idea: agents talk with the A2A protocol

The Orchestrator and the SRE agent are separate agents. They talk with the Agent2Agent (A2A) protocol: JSON-RPC over HTTP, with streamed progress.

Each A2A agent publishes an agent card at /.well-known/agent-card.json. The card says who the agent is, where to send requests, and which skills it has. A client reads the card, then names a skill in its request.

flowchart LR
    O["🤖 Orchestrator"] -->|1 · GET /.well-known/agent-card.json| C["📇 agent card<br/>skills: list_incidents,<br/>diagnose_incident, …"]
    O -->|2 · message + metadata skill=…| S["🩺 SRE agent"]
    S -->|3 · progress, then the result| O

sre_agent/src/sre_agent/a2a_agent.py builds the card, and to_a2a() from ADK serves the agent. _skill_run sends each request to the pipeline of its skill.

At this time, the card has two skills, list_incidents and diagnose_incident. A request for write_post_mortem fails with Unknown skill.

Your task

  1. Find the TODOs:

    git grep -n "TODO(step-3)"

  2. In _skill_run, run the WRITE_POST_MORTEM skill with run_post_mortem(...).

  3. In build_agent_card, add an AgentSkill with id=WRITE_POST_MORTEM. Use the two skills above it as examples. The description is for other agents and their models: say what the skill does, and which request metadata it reads (skill, project_id, trace_id).

Run it

  1. Do this command:

    uv run workshop/check.py 3

  2. Start the SRE agent alone:

    uv run uvicorn sre_agent.main:app --port 8081

  3. Open http://localhost:8081/.well-known/agent-card.json in your browser. Find your skill.

  4. Press Ctrl+C to stop the SRE agent.

The chat cannot use the skill yet: the Orchestrator does not call the SRE agent until step 4.

Discuss

  • Why is the description so long? Another agent's model reads it to decide when to use the skill. Write it like a tool docstring.
  • Why one skill for each type of answer, and not one skill with a parameter? A client can see the skills on the card, and a policy can allow or deny each one (step 4).

Stuck? Run uv run workshop/step.py hint 3, or ask the workshop coach in Antigravity. To apply the solution: uv run workshop/step.py solve 3.

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/a2a_agent.py, in _skill_run(). The TODO says:

# TODO(step-3): Run the WRITE_POST_MORTEM skill: return
#   run_post_mortem(prompt=prompt, project_id=project_id, trace_id=trace_id, ui=ui).

File: sre_agent/src/sre_agent/a2a_agent.py, in build_agent_card(). The TODO says:

# TODO(step-3): Publish the post-mortem skill: an AgentSkill with id=WRITE_POST_MORTEM,
#   a name, a description (what it does, and its request metadata: skill, project_id,
#   trace_id), tags, examples and output_modes. Use the two skills above as examples.
Solution 
--- a/sre_agent/src/sre_agent/a2a_agent.py
+++ b/sre_agent/src/sre_agent/a2a_agent.py
@@ -53,8 +53,8 @@
     ui = a2ui_surfaces.client_renders_sre_catalog(metadata)
     if skill == LIST_INCIDENTS:
         return run_list_incidents(project_id=project_id, ui=ui)
-    # TODO(step-3): Run the WRITE_POST_MORTEM skill: return
-    #   run_post_mortem(prompt=prompt, project_id=project_id, trace_id=trace_id, ui=ui).
+    if skill == WRITE_POST_MORTEM:
+        return run_post_mortem(prompt=prompt, project_id=project_id, trace_id=trace_id, ui=ui)
     if skill not in ("", DIAGNOSE_INCIDENT):
         raise ValueError(
             f"Unknown skill {skill!r}: expected one of {LIST_INCIDENTS}, {DIAGNOSE_INCIDENT}, {WRITE_POST_MORTEM}"
@@ -174,9 +174,19 @@
                 ],
                 output_modes=["text/markdown"],
             ),
-            # TODO(step-3): Publish the post-mortem skill: an AgentSkill with id=WRITE_POST_MORTEM,
-            #   a name, a description (what it does, and its request metadata: skill, project_id,
-            #   trace_id), tags, examples and output_modes. Use the two skills above as examples.
+            AgentSkill(
+                id=WRITE_POST_MORTEM,
+                name="Write a post-mortem",
+                description=(
+                    "Writes the incident post-mortem of one trace from its spans and logs: overview, timeline, "
+                    "root cause and next steps. With a Gemini key, adds AI-generated analyst notes. Request "
+                    'metadata: skill="write_post_mortem", project_id, trace_id (optional: defaults to the most '
+                    "important recent incident)."
+                ),
+                tags=["sre", "post-mortem"],
+                examples=["Write the post-mortem for trace 1c65bf87e4be434ea6d6d7edc1ef8c97."],
+                output_modes=["text/markdown", "application/json"],
+            ),
         ],
     )

To apply it to your code: uv run workshop/step.py solve 3.