A course, written by a small team of agents

What the afternoon lab was

Calgary Innovation Week, 4 November 2026. The lab builds a course creator as separate services instead of one prompt that tries to do every job. All 10 codelab steps passed. Cleanup afterwards left no Cloud Run services.

The talk's picture is a team. The researcher has web search. The judge can send the work back. The writer formats the course. The orchestrator keeps the shared state. A sequential pipeline runs the research loop first and the writer second. Agent-to-agent calls let the specialists live apart.

The public codelab is Building a multi-agent system. The starter sample is prai-roadshow-lab-1-starter in GoogleCloudPlatform/devrel-demos, Apache-2.0, Copyright 2022 Google LLC.

The afternoon reused the morning workshop's cloud project. Sign-in was Vertex through Application Default Credentials. Cloud Run used region us-west1. The model location was global. The lab turned on Cloud Run, Artifact Registry, Cloud Build, Vertex AI, and Compute. This page leaves out the project id, the project number, service URLs, and machine paths.

How a course gets built

The same flowchart as the lab, drawn as HTML so it reflows. Solid steps are control flow inside the orchestrator. The A2A list is how that orchestrator actually reaches the three specialists.

What each agent does

Three specialists do the work. The orchestrator does not search, grade, or write. It only decides who runs next. Every model agent used gemini-3-flash-preview.

What A2A added

The morning workshop kept weather in-process and sent only currency out over A2A. This lab sends every specialist out. The orchestrator does not import their Python modules.

What the run verified

All 10 codelab steps passed. These are the checks that were written down. Outputs are summarised.

1. 1

   Researcher, alone

   Prompt: "Find the population of Tokyo in 2020." The reply cited the 2020 Japan Census.

   Tokyo Metropolis

   14,047,594

   Versus 2015

   +3.9%

   Density

   6,402 persons per km²

   Daytime population

   16.752 million

   23 special wards

   9,733,276

   Greater Tokyo

   about 36,914,000

2. 2

   Judge, alone

   Prompt: "Topic: Tokyo. Findings: Tokyo is a city." The judge failed it. The object returned:

   {"status": "fail", "feedback": "The findings are extremely insufficient for the topic 'Tokyo'. They lack critical information such as its status as the capital of Japan, population statistics, major landmarks, economic importance, and cultural significance."}

3. 3

   Five processes on the laptop

   Ports 8000 through 8004. The three specialist cards matched the skills above, with JSON-RPC and streaming.

4. 4

   Cloud Run, then deleted

   Five services in us-west1, labelled for the lab: researcher, judge, content-builder, orchestrator, and course-creator. They were removed in cleanup. This page does not publish their URLs or revision ids.

5. 5

   End to end: history of coffee

   Prompt: "Create a course about the history of coffee." The stream reported research, then a judge pass, then the writer, then the course. The judge's pass:

   {"status": "pass", "feedback": "The research findings provide a comprehensive and well-structured course syllabus covering the history of coffee from its origins in Ethiopia to the modern third-wave movement, including cultural, economic, and political contexts."}

   The course title was "The Black Gold: A Global History of Coffee." The sections were:

   1. The myth and mysticism of discovery (800 CE to the 1400s).
   2. The Ottoman bloom and the birth of the coffeehouse (1500s to the 1600s).
   3. Europe's penny universities and the Enlightenment (1600s to the 1700s).
   4. Colonialism and slavery (1700s to the 1800s).
   5. Machines and the first wave (1800s to the 1960s).
   6. Modern waves, from the second wave to specialty (1960s to the present).
   7. Practical applications and further reading.

Two things that broke, and the fixes

1. 1

   Google Search was not a local Python tool

   The error was ValueError: Tool 'google_search' not found. In this ADK version, google_search turns on server-side Gemini grounding. It is not a function registered in the local tool map. An instruction that said to use the google_search tool, with that name in backticks, made gemini-3-flash-preview emit a client-side function call of that name. The working sentence is "Use Google Search to find relevant information."

2. 2

   Three deploys at once hit a Cloud Build limit

   The error was 429 RESOURCE_EXHAUSTED while the deploy command polled Cloud Build. Three gcloud run deploy --source commands at the same time meant three pollers, about 20 status reads a minute each, against a quota of 60 reads a minute. Deploying at most two services at a time stayed under that limit.

How I'd use this pattern

A high-variance draft should not be the thing that publishes. Put a short loop in front of the publish step.

Cleanup confirmed

Nothing from this lab was left running. The five Cloud Run services were deleted in us-west1: researcher, judge, content-builder, orchestrator, and course-creator. The source-deploy image repository was deleted too, which freed about 755 MB.

A list filtered to the lab label, and a list of Cloud Run services across regions, both came back as:

Listed 0 items.

Stack that ran

Resolved from the lockfile on CPython 3.13.18. Every LLM agent used gemini-3-flash-preview. The A2A cards reported protocol 0.3.0. The installed library was a2a-sdk 0.3.22.

• google-adk 1.28.0
• google-genai 1.68.0
• a2a-sdk 0.3.22
• fastapi 0.124.4
• uvicorn 0.40.0
• pydantic 2.12.5
• google-cloud-aiplatform 1.132.0
• CPython 3.13.18

Code from the sample

The judge schema and the escalation checker below are the versions that ran on 4 November 2026. They derive from Google's Apache-2.0 sample prai-roadshow-lab-1-starter (GoogleCloudPlatform/devrel-demos, Copyright 2022 Google LLC). The starter leaves those two pieces as exercises. This is the completed lab code, under that licence. The notice is also in notice-lab1.txt.

Copyright 2022 Google LLC

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

Judge schema

From agents/judge/agent.py. MODEL is gemini-3-flash-preview.

class JudgeFeedback(BaseModel):
    """Structured feedback from the Judge agent."""
    status: Literal["pass", "fail"] = Field(
        description="Whether the research is sufficient ('pass') or needs more work ('fail')."
    )
    feedback: str = Field(
        description="Detailed feedback on what is missing. If 'pass', a brief confirmation."
    )


judge = Agent(
    name="judge",
    model=MODEL,
    description="Evaluates research findings for completeness and accuracy.",
    instruction="""
    You are a strict editor.
    Evaluate the 'research_findings' against the user's original request.
    If the findings are missing key info, return status='fail' and specific feedback on what is missing.
    If the findings are comprehensive and accurate, return status='pass'.
    """,
    output_schema=JudgeFeedback,
    disallow_transfer_to_parent=True,
    disallow_transfer_to_peers=True,
)

Escalation checker

From agents/orchestrator/agent.py. In that file, researcher, judge, and content builder are remote A2A agents. The checker, the loop, and the sequential agent are local.

class EscalationChecker(BaseAgent):
    """Checks the judge's feedback and escalates (breaks the loop) if it passed."""

    async def _run_async_impl(
        self, ctx: InvocationContext
    ) -> AsyncGenerator[Event, None]:
        feedback = ctx.session.state.get("judge_feedback")
        print(f"[EscalationChecker] Feedback: {feedback}")

        is_pass = False
        if isinstance(feedback, dict) and feedback.get("status") == "pass":
            is_pass = True
        elif isinstance(feedback, str) and '"status": "pass"' in feedback:
            is_pass = True

        if is_pass:
            yield Event(author=self.name, actions=EventActions(escalate=True))
        else:
            yield Event(author=self.name)


escalation_checker = EscalationChecker(name="escalation_checker")

research_loop = LoopAgent(
    name="research_loop",
    description="Iteratively researches and judges until quality standards are met.",
    sub_agents=[researcher, judge, escalation_checker],
    max_iterations=3,
)

root_agent = SequentialAgent(
    name="course_creation_pipeline",
    description="A pipeline that researches a topic and then builds a course from it.",
    sub_agents=[research_loop, content_builder],
)