Create and configure a Sandbox agent (BETA)
An agent is a reusable profile you set up once and then start as many times as you need. It defines the persona and environment of your Sandbox conversations: a name, a runtime, a model, and instructions. Creating agents is part of the Sandbox Agent BETA feature (overview).
Create an agent
From the Sandbox section of the dashboard, open the agents table and click New agent. You can create, edit, and delete your agents at any time. The limit is 20 agents per account.
Configuration fields
- Name — the agent's label, shown in the table and in your conversations.
- Avatar / color (optional) — an avatar and color to spot the agent at a glance in the list.
- Runtime — the sandbox image the agent runs in: Python or Node. Pick it based on the language the agent needs to work with. The runtime determines which tools and libraries are available by default in the sandbox.
- Model — the LLM that drives the agent (see models). Coding/agentic models are recommended for a Sandbox agent.
- Instructions — the agent's system prompt / persona: what it is, what it should do, its rules. Maximum 4000 characters. This is the most important field: be specific about the goal, the expected output format, and which tools to favor.
Writing good instructions
- Describe the task and the expected result, not just the tone.
- Say when to publish a downloadable artifact.
- Remind the agent that its sandbox is ephemeral (~10 min): anything worth keeping must be published or checkpointed.
The agents table
The table lists your agents with each one's sandbox status. It provides search and filters by runtime and status so you can find an agent quickly. From a row you can start a conversation or edit/delete the agent.
Next steps
- Conversations and turns — start a conversation from the agent.
- Agent tools — what the agent can do.