AgentKit quickstart

Run a small greeting workflow locally, then use the same files on Adios. It produces a structured greeting and a greeting.txt artifact without accessing an external service.

Get the example

Download hello-workflow.zip and extract it. The archive includes instructions, JSON schemas, a workflow, and recorded test data.

hello-workflow/
  agentkit.yaml
  instructions/greeter.md
  schemas/input.json
  schemas/output.json
  tests/input.json
  tests/recording.json

Install the Adios CLI to validate or upload the example:

adios agent validate ./hello-workflow

The standalone agentkit development CLI is built from an AgentKit source checkout:

make build
./dist/agentkit help

Run the following command from that checkout, replacing the example path with the directory you extracted. This recorded run needs neither login nor a model key.

./dist/agentkit workflow --workflow hello \
  --input ./hello-workflow/tests/input.json \
  --recording ./hello-workflow/tests/recording.json \
  --state ./.agentkit/hello-recorded --run-id hello-recorded \
  ./hello-workflow

The result should have status completed, with the greet step output:

{"message":"Hello, Adios!"}

The greeting step saves greeting.txt containing that message. Reusing the state directory and run ID resumes the saved run. Choose a new run ID to test changed input or instructions.

Use a real model

Replace --recording with your model server's URL and model name:

./dist/agentkit workflow --workflow hello \
  --model-url http://localhost:1234/v1 --model YOUR_MODEL \
  --input ./hello-workflow/tests/input.json \
  --state ./.agentkit/hello-model --run-id hello-model \
  ./hello-workflow

The server must implement the OpenAI-compatible chat completions interface. For an authenticated endpoint, add --model-key-env AGENT_MODEL_KEY after setting that variable locally. The flag names the variable, not its value.

Run the same kit on Adios

Authenticate and select your test team using the CLI setup guide.

adios agent create --dir ./hello-workflow --name "Hello workflow"
adios agent run AGENT_ID --test --workflow hello --input '{"name":"Adios"}'

Use the agent ID returned by creation. Compare the greeting and artifact with the local result in the run's Activity view. Hosted test runs may make real model calls and consume usage. Continue with the hosted guide for versions, connections, approvals, and normal runs.