Getting started¶
Launch the Agno AgentOS API, point it at a model you are entitled to use, and check the CSV calculation after a service restart.
You need: an AWS account that can subscribe and launch EC2, an instance
profile with AmazonSSMManagedInstanceCore, and an OpenAI-compatible model
that supports tool calling and non-streaming chat completions. Session
Manager needs the Session Manager plugin on your workstation. SSH is optional.
1. Subscribe¶
Open the Agno AgentOS API listing, choose View purchase options, review the terms, and subscribe.
2. Launch the instance¶
On the EC2 launch page:
- Choose t3.medium (4 GiB of memory) or a larger listed type.
- Use an encrypted gp3 root volume of at least 24 GiB.
- Require IMDSv2.
- Attach an instance profile with
AmazonSSMManagedInstanceCore. No key pair and no inbound rule are required for Session Manager. - Leave application ports closed. The API listens only on
127.0.0.1:7777.
3. Connect¶
Session Manager starts as ssm-user. Wait for first boot, then check health.
The service can answer health before a model is configured.
First boot creates a unique API token. Settings, including the provider key
once you add one, are in /home/ubuntu/.config/sushi-agno/settings.json and
must stay mode 0600.
4. Configure a model¶
In a terminal that is not being recorded:
Enter the model id and an OpenAI-compatible base URL. For OpenAI the public
guide uses https://api.openai.com/v1 and a model your account can call.
The key prompt does not echo. A model server on this instance can use a
loopback HTTP URL. This image does not install or download one. Remote
providers need HTTPS, tool calling, and non-streaming chat completions.
5. Run the CSV task¶
The demo writes input.csv with the values 5, 17, and 20, calls the
calculation tool, checks that the sum is 42 and the row count is 3, and
reads the stored session. A successful run prints SUSHI_CUSTOMER_DEMO_PASSED.
Task files stay under /var/lib/agno/workspace/. Network and provider
failures can still incur provider charges.
6. Read it again after a restart¶
The public guide uses the directory the demo printed:
task_directory=$(sudo -iu ubuntu sushi-agno demo)
sudo -iu ubuntu python3 -I -B /opt/agno/readback.py "$task_directory"
sudo systemctl restart sushi-agno
sudo -iu ubuntu sushi-agno session "$task_directory"
sudo -iu ubuntu python3 -I -B /opt/agno/readback.py "$task_directory" --after-restart
sushi-agno session writes its restart readback once and refuses to
overwrite it. These checks cover an idle restart. They do not promise
exactly-once execution if the process stops during a model or tool call.
Last updated November 11, 2026 · Suggest an improvement