Agents
Create and configure an agent
An agent is a name, a type, an optional set of tools, and a block of configuration that holds its instructions. I built this form, and the thing that trips people up is the last part: there is no separate “instructions” box. The instructions live in the Configuration JSON, under system_prompt.
This page is the field reference. If you want to be walked through a whole build instead — one agent, one tool, a real run read line by line, and the step where it gets something wrong — use Build a support agent that answers a refund ticket. It covers the same form in the order you actually fill it in.
And if you only want to see one work, the Dashboard has a one-click starter agent — see Get started in 3 steps.
Create it
- Go to Agents → Create agent.
- Name it something you will recognise in a run list six weeks from now.
- Pick a type. It labels the agent and drives its icon; it does not change how the model behaves.
- Describe it, for your own benefit. The description is not sent to the model.
- Tick any tools it may use. An agent gets nothing it was not explicitly given.
- Write its instructions in the Configuration (JSON) box, as shown below.
- Save.
The Configuration JSON, which is the part that matters
This box is where the agent actually gets its behaviour. A minimal, working example:
| Key | What it does | Default if you omit it |
|---|---|---|
system_prompt | The agent’s instructions. This is the one that decides what it does | No system message is sent at all — the model sees only the input |
model | Which model to run | The platform default |
temperature | How much the output varies between runs | 0.7 |
max_tokens | Cap on the length of the reply | The provider’s own default |
response_format | Set to json to force a single JSON object back | Plain text |
An agent saved with an empty Configuration will run and answer, but it answers as a general assistant, because nothing told it what its job is. That is the most common reason a new agent “ignores” what you wanted.
Pick a type
Seven are available, and the choice is organisational rather than functional: CUSTOMER_SUPPORT, DATA_PROCESSING, WORKFLOW_AUTOMATION, CONTENT_GENERATION, ANALYSIS, INTEGRATION, CUSTOM. Use the one you will search for later.
Activate and run it
Agents are created Inactive on purpose, so nothing runs by accident and nothing spends tokens before you meant it to. Toggle it to Active, click Run, give it an input, and the run appears in Executions with its steps, tokens and cost.
A first run is cheap. Source: the worked example shows a support-ticket job at 8.4 seconds and $0.0144, and a run that failed because its API was down at $0.0024 — both costed by the same pricing code the product bills with, readable without an account.
None of the settings above change what a run costs. Tokens do. A worked support example comes to $0.0144 at our real rates, and a real logged support run that also escalated to a human came to $0.0204, so a longer prompt and a longer answer are the two things that move the number.
Every change is kept
Saving an edit writes a version-history entry with a full snapshot of the agent and who changed it, tagged as a create, an update or a rollback. So you can see what an agent looked like before a change, and roll back to it, which matters more than it sounds once an agent has been running for a month.
When it does not behave
It ignores your instructions
Check the Configuration box is valid JSON and that the key is spelled system_prompt. A typo means no system message is sent and the agent falls back to being a general assistant — it will not warn you.
Nothing runs
The agent is Inactive. That is the default and it is deliberate.
You need its output in a workflow
Turn on structured JSON output, either with the checkbox on the form or "response_format": "json" in the configuration. The agent is then instructed to return a single JSON object, and later steps and conditions can read its fields by name.
Next
- Let it act on real systems — see Give agents tools & integrations.
- Chain it with other agents — see Build a multi-agent workflow.