Sandbox Agent conversations and turns (BETA)

Once you've configured an agent, you talk to it through conversations made of turns. Each turn spins up an ephemeral Docker sandbox where the agent does its work, then streams a response back to you. This page covers a turn's lifecycle in the Sandbox Agent BETA feature (overview).

Start a conversation

From the agents table, pick an agent and start a conversation. A conversation holds the history of your exchanges with that agent. You can rename or delete it at any time.

Run a turn

Send a message: that's a turn. On every turn:

• Velqa spins up a fresh sandbox with the agent's runtime (Python or Node).
• The sandbox has a lifetime of roughly 10 minutes before it expires.
• The response streams as it's generated. The UI shows the agent's tool calls (see tools), the elapsed time, and the sandbox expiry.

This lets you follow, live, what the agent does: which files it writes, which commands it runs, which searches it performs.

Execution limits

To keep sandboxes isolated and available during the beta, turns are subject to caps:

• Expiry — a sandbox lives about 10 minutes; after that it's stopped.
• Concurrency — only a limited number of sandboxes can run at once.
• Hourly rate — a cap on turns per hour.

If a cap is hit, the UI tells you; retry a bit later or wait for a running turn to finish.

Stop a turn

You can stop a running turn: this halts generation and the associated sandbox. Handy if the agent heads the wrong way or you already got what you needed — it also frees up concurrency.

The agents table and its filters

The agents table shows each agent's current sandbox status. Search and filters by runtime and status help you quickly spot agents that are active, pending, or idle.

Next steps

• Agent tools — the detail behind each tool call you see during a turn.
• Workspace and artifacts — give files to the agent and collect its deliverables.