Managed agents

A managed agent is an agent Runboard runs for you. It can run as a Runboard Sandbox — a fresh cloud workspace for a single task — or as a Runboard Pod, a dedicated cloud machine with Codex or Claude Code installed (signed in with your own subscription for that Pod). There is no local machine to maintain: assign it an issue and it gets to work — clones the repository, makes the changes, and reports progress, a summary, a git diff, and (optionally) a pull request straight back to the issue.

It is still a normal agent: assignment, comments, and activity all work the same way.

Under Settings → Agents, use Run an agent in Runboard to set up a Runboard Sandbox agent, then configure:

Model

The model that drives the run — search by name or provider.

Repository

The repo the agent clones to work in. Leave it empty to default to the issue's project repository, or pick No repository for research-only runs. Connect repositories under Integrations.

Instructions

Project context, conventions, or guard rails applied to every run.

Start automatically when assigned

When on, assigning an issue to this agent kicks off a run with no extra click. On by default.

Open a pull request

Push a branch and open a PR when there are changes. On by default.

Max duration & vCPUs

How long a run may take (up to 44 minutes) and how much compute it gets (1–8 vCPUs). These also affect cost — see Billing.

For a dedicated machine instead, use Run an agent on a managed host to create a Runboard Pod. Choose the agent to install, machine size, repository behavior, and auto-start setting. A Pod signs in with your own saved Codex or Claude Code subscription; Runboard encrypts it and uses it only when dispatching a run. The Pod keeps its workspace between tasks and can stop itself when idle.

There are three ways to start a run, and one of them needs no clicks at all:

1. 1Assign an issue to a managed agent that has auto-start on — the run begins immediately
2. 2Use the Run agent control on an issue to launch a specific managed agent
3. 3Call the start_agent_run MCP tool, so a remote agent can dispatch a managed run too

Runs can also be ad-hoc — started without an issue, for a one-off task.

While it works, a run streams its output live. When it finishes, it leaves behind everything you need to review the work:

• A summary of what it did and a streaming log
• A git diff — files changed, with additions and deletions
• A branch and, if enabled and there were changes, a pull request
• A comment on the issue and, where appropriate, a status transition
• Usage — Sandbox runs show tokens, model cost, and compute; a Pod shows run status while its machine uptime and storage are billed through credits

Runs lists every run in the workspace with its live status. Open one for the full detail: a streaming log viewer, the summary, the exact task prompt, the diff, the pull request link, the model and repository, and a usage panel. A run moves through these states:

Cancel an in-progress run from its page or with the cancel_agent_run tool. Each issue also has an Agent runs section showing its own run history.

Runboard Sandbox and Runboard Pod require a managed-agent plan (Pro and up). AI credits pay for Sandbox runs plus Pod uptime and storage; a Pod's Codex or Claude Code model usage is billed to your own subscription for that Pod. Each workspace also has caps for agents running in parallel and for Pods. See Billing for credits, plans, and limits.