Sandboxes for your agents
Each agent gets an isolated computer with full access to your workspace: it can analyze files, run code, and coordinate with other agents.
Generated weekly executive summary from data-analyzer outputs. 3 files written to outputs/.
Reviewing auth middleware changes across 14 files. Reading previous session history for context.
Powered by Runwork AI What an agent sandbox is
An agent sandbox is a computer an AI agent gets to itself: its own filesystem, its own runtime, its own network access, walled off from everything else. The agent can install things, write files, run code and read the results, the way a person would at a terminal, and nothing it does reaches anything outside that boundary.
The isolation is the point. An agent that can only call pre-approved tools is limited to what someone anticipated. An agent with its own machine can do work nobody wrote a tool for. The sandbox is what makes the second one safe enough to allow.
In practice that is the work that does not fit in a chat reply: analysing a spreadsheet properly, processing files in bulk, research that takes an hour, anything long enough that you would rather it ran somewhere else while you got on.
What Agent Sandbox gives you
Delegate Any Task
Agents hand off complex work to a coding agent running in an isolated sandbox. Data analysis, report generation, code review, file processing, multi-step research. If it can be done at a computer, your agent can do it.
Full Computing Environment
Not a constrained tool call. A real computer. File system, web access, MCP tools, workspace integrations. Your agents work with the same power a developer has in their terminal.
Cross-Agent Coordination
Agents read each other's outputs and continue previous sessions. No complex protocols needed. Agents coordinate through the filesystem, the simplest interface there is.
Session Persistence
Resume work where you left off. Pass a session ID and the agent picks up with full context of previous work. Long-running analysis, iterative refinement, multi-day projects.
Cost Controls
Three-layer budget system keeps spending predictable. Platform credits, agent-level limits, and per-task overrides. Budget warnings at 90% let agents wrap up gracefully. Hard stops at 100%.
Programmatic or Built-in
Every agent automatically gets the run_agent_task tool. Or call runAgentTask() in code for full control over inputs, outputs, and execution limits. Both paths, same power.
Execution Stream
Watch sandbox agents work in real time from the dashboard. Every tool call, text output, and error rendered with timestamps. Expand any execution to see the full event log.
Token and Cost Tracking
Every execution tracks input tokens, output tokens, and USD cost. See cost per run in the execution history. Budget warnings at 90% let agents wrap up gracefully before the hard stop.
Access from anywhere
Desktop & Web
Codex DEFAULT 
Claude Code 
Cursor Watch execution streams in real time from the Agents Dashboard. See tool calls, costs, token usage, and output files. Retry failed runs with one click.
CLI
Why Agent Sandbox matters
- No infrastructure to manage. Sandboxes are provisioned automatically.
- Workspace credentials flow to the sandbox. No API key juggling.
- It runs a real coding agent at a real terminal, not a cut-down version built for a product demo.
- Agents coordinate through the filesystem. Simple, no protocol overhead.
- You don't need a separate app to run an agent. The workspace computer is already an app.
How Agent Sandbox works
The Agent Sandbox gives your AI agents their own computing environment. When an agent needs to do complex work, it delegates the task to a coding agent running inside an isolated sandbox with full access to your workspace.
The framework agent decides what needs doing and calls run_agent_task with a prompt, input files, and output patterns. The platform provisions a sandbox session in the workspace computer. The agent runs with full workspace context: your skills, MCP servers, integrations, and data. Results flow back to the orchestrating agent.
Agents share work through the filesystem. Every agent session lives in /workspace/agents/ where other agents can read previous outputs and continue where someone left off. No inter-agent messaging protocol needed. Just files.
You don't need to build a separate app just to run an agent. The workspace computer is itself an app. Write agents directly against your workspace for single-purpose tasks: a new schedule, a one-off analysis, a recurring report. The CLI works the same way locally, so the experience is consistent everywhere.
Cost controls work in three layers. Your platform credit balance sets the ceiling. Agent definitions set per-task budgets, turn limits, and timeouts. Individual runAgentTask() calls can tighten these further. At 90% of budget, the agent gets a warning to wrap up. At 100%, it stops gracefully and returns whatever results it has.
The Execution Stream shows exactly what your sandbox agents are doing. From the Agents Dashboard, expand any execution to see the full event log: every tool call with its input parameters, text responses, errors, and status changes, all with timestamps. The stream auto-updates while agents are running. Each execution also tracks token usage (input and output) and cost in USD, so you see exactly what every run costs. Failed or partial executions can be retried with one click. From the terminal, runwork agents execution <sessionId> shows the same stream.
Frequently Asked Questions
What is an agent sandbox?
How is that different from giving an agent tools?
What can agents do in the sandbox?
How do agents share data with each other?
Do I need to create an app to use agent sandboxes?
How are costs controlled?
Which AI actually runs in the sandbox?
Can I run coding agents in isolation from each other?
How do I sandbox agents running inside my own team's tools?
Use Cases
Related Features
See How Teams Use Agent Sandbox
Ready to try Agent Sandbox?
One shared cloud under the AI tools your team already uses.