Great question — in Prefect 3.x the ECS task definition image (the container that actually runs your flow) should be a runtime image that includes: - Python (matching your code) - Prefect 3.x - Any libraries your flow needs (plus any system libs) There isn’t a special “worker run” image to use. Most teams either: - Build a slim custom image from Python and install Prefect + deps, or - Start from the official Prefect base image and add their dependencies Minimal custom image example:

# Start from a slim Python base
FROM python:3.11-slim

# Install Prefect and AWS extras (if your flow uses AWS libs)
ENV PIP_NO_CACHE_DIR=1
RUN pip install --upgrade pip \
 && pip install "prefect>=3,<4" "prefect-aws"  # or: pip install 'prefect[aws]'

# Optionally: install your project requirements
# COPY requirements.txt .
# RUN pip install -r requirements.txt

# Optionally: add your code if you bake flows into the image
# COPY . /opt/project
# ENV PYTHONPATH=/opt/project

Where to set the image: - Default for the pool: In your ECS work pool’s base job template (UI or API), set the container definition’s

image

to your ECR/Docker image. - Per deployment: Provide an image when deploying your flow so each deployment can pin its own runtime image. In code this typically looks like:

from prefect import flow

@flow
def my_flow():
    ...

if __name__ == "__main__":
    (
        my_flow
        .from_source("<https://github.com/your-org/your-repo>", "path/to/flow.py:my_flow")
        .deploy(
            name="prod",
            work_pool_name="my-ecs-pool",
            image="<http://123456789012.dkr.ecr.us-east-1.amazonaws.com/myflow:2025-09-24|123456789012.dkr.ecr.us-east-1.amazonaws.com/myflow:2025-09-24>"
        )
    )

Recommendations and notes: - Prefer pinning a specific tag (not

latest

) to make runs reproducible. - Make sure your ECS cluster/role can pull the image (ECR permissions). - If you interact with AWS services from the flow, include the needed libs (boto3, s3fs, etc.). - The worker image and the flow-run image are separate — the worker only launches ECS tasks; your flow code runs inside the image you specify above. Docs to dig deeper: - Prefect AWS ECS guide - How to store and ship flow code in deployments If you share your Python version and any heavy dependencies, I can sketch a Dockerfile tailored to your stack.