Dockerfile, agentos.yaml, code) you can zip and deploy as it is, then read to see how a feature is done.
Agents that call a model name one in
config.models (for example main: anthropic/claude-sonnet-4-5). Before you upload one, change it to a model from your environment’s Models page: <provider slug>/<model ID>. See Models.Without an SDK (plain HTTP)
These use only the Python standard library, so they show every platform call as raw HTTP. Read them if you are writing an agent in a language without an SDK.hello-agent
The smallest agent that follows the contract: one capability,
echo, that returns its input. Used in the Quickstart.llm-agent
ask: answers a question with one model call through the gateway, using the run token, and posts an event to the run’s timeline. See Models.review-agent
pay_invoices asks a person to approve, edit or reject payments and keeps its progress in run state; chat keeps a conversation’s history in session state; split starts one child run per invoice. See Human approvals.stream-agent
chat streams a model’s answer as a Vercel AI SDK UI message stream and reports the run’s result itself; confirm asks a person mid-stream. Start it with ?passthrough=true and a chat UI reads the stream as it is. See Streaming.With the SDKs
py-sdk-agent
review-agent and llm-agent rewritten with the Python SDK: approvals, run, session and agent state (with compare-and-set), child runs, a model call, an event, a file upload, a streamed answer (
tell) and notes kept per end user in memory (notes).ts-sdk-agent
The same capabilities with the TypeScript SDK, on Node 24.
file-agent
inspect takes any file (and an optional second one), downloads it with the run token and returns its type, size, SHA-256 and the start of its text. No model call. Good for trying file inputs on the Run tab. See Files and events.ocr-agent
extract reads a (fake) Turkish company document image with a vision model and returns its fields. The model is config.models.reader, so an experiment can compare vision models. Comes with a script that generates sample documents and their expected answers. See Evals.math-agent
solve answers a math question with a single number, so a reference evaluator can compare it as a number. A minimal agent for trying evals and model comparisons.Framework examples
The SDKs also include one small agent per framework adapter, each with a payment that waits for a person:File inputs
file-agent and ocr-agent declare their file inputs as {type: string, format: file} in the input schema (ocr-agent adds contentMediaType: "image/*"). The panel’s Run tab then shows an upload for the field, and the value the agent receives is a file id it downloads with the run token.
Conformance cases
llm-agent, review-agent, stream-agent, py-sdk-agent and ts-sdk-agent each have a conformance.yaml next to their agentos.yaml: a good starting point for your own. See Conformance kit.