Your first agent

Build a small question-answering agent first. This proves that your environment, model connection, publication, and run access work before you add integrations.

1. Check your environment and model

Choose the intended organization, project, and environment in the console. Open Connections and find a model provider profile you can use. If none exists, an owner or administrator can register one using the Models and connections guide.

A profile being listed is not the same as being ready. Its connection must be approved, the environment must allow it, and the person starting the run must have the required access. Resolve the setup messages in Connections before your first run.

2. Create a draft

1. Open Agents and choose New agent.
2. Choose the Hello world starter.
3. Enter a recognizable name, such as Request assistant.
4. Select your model under Model.
5. Choose Create draft.

The starter takes a question and returns text. It already has an objective, an instruction, and the required output-schema criterion. It does not need attached tools, knowledge, or specialists.

3. Inspect the behavior

In Overview, read the Objective and Instructions. For this first run, keep the supplied objective: answer the question in one sentence. An objective describes the outcome; instructions describe how to approach it.

Open Input & output. Keep A message as the input and A written answer as the output. Run form preview shows what someone starting the agent will fill in.

If you change anything, choose Save draft. Optional Run limits live in Overview. Leave them empty while you are learning, or set deliberate ceilings for steps, model turns, tool calls, and duration. An empty field means the agent defines no cap for that field; other environment and model limits still apply.

4. Check input, then publish

Choose Test…, enter a short question, and choose Check input. For an unpublished draft, this checks the input against its schema; it does not call the model. You still need to publish before you can run the agent.

Choose Publish…. Review the model and the checks in the dialog, resolve any publication blockers, and choose Publish v1. With this starter there are no knowledge collections to bind and no queue is required.

The Before a run can use it and Would v1 run? checks are useful even when publication is allowed. A published version can still be blocked for a particular person. See Publishing and versions for the distinction.

5. Start the published version

Open the published version from Versions and start a run, or use the run launcher to choose the agent and v1. Enter a question such as:

What is the purpose of an acceptance criterion?

Submit the run and open its workspace. Check that it reaches a successful outcome and returns one sentence. Read the saved answer and any acceptance results. Streaming text, if enabled later, is provisional until the run finishes and its result is assessed.

6. Try a change deliberately

Return to the draft. Add an instruction to use plain language, save it, and publish the next version. Run the same question against that version and compare the outcomes. The first version and its earlier run remain unchanged.

Test more than a happy path: an empty or invalid input, a question missing important information, and a question outside the agent's intended purpose. Use those results to improve the objective, instructions, and output contract.

What you have now

You have a reusable agent definition, a published version, and a recorded run. Next, use Configure an agent to choose the settings your task needs. Add knowledge for reference material, tools and skills for capabilities, or a workflow for a process with an explicit review step.

If creation, publishing, or the run is blocked, start with the exact message and the agent's Access section. The Troubleshooting guide maps common symptoms to the right screen.