Agents

What it is

An Agent is a saved, reusable bundle of a prompt plus the model setup to run it: prompt + provider + model. Where the Playground is a scratchpad, an Agent is a named configuration you keep and run again.

Why it matters

In practice you don't just have prompt text — you have "this prompt, run on this provider, with this model". An Agent captures that whole unit once, so anyone on the team can run it consistently without re-entering the setup.

Anatomy of an agent

Field	Description
key	Stable identifier, e.g. support.agent
name	Human-readable label
description	What this agent is for
prompt	The prompt text, with placeholders
provider	mock / openai / claude / ollama / gemini
model	Model name; leave empty for the provider default

Using it

1. Open the Agents page and click New Agent.
2. Fill in the key, choose a provider and model, and write the prompt.
3. Click Save.
4. On the agent's page, fill in any detected , supply an API key if the provider needs one, and click Run Agent.

The result shows the rendered prompt and the model output, just like the Playground — but the configuration is saved for next time.

Behind the scenes

Running an agent calls POST /api/agents/:id/run. Each run is recorded for Observability.

Agent vs. Workflow

• An Agent is a single prompt-and-model step.
• A Workflow chains multiple steps, feeding each step's output into the next.

Next steps

• Workflows — orchestrate multiple steps.
• Observability — review agent runs.