Your AI agents now need budget approval and a sandbox

What it does

Magic is an open-source TypeScript platform that bundles generalist AI agents, a workflow engine, instant messaging, and collaborative editing into one system. Built as an enterprise-hardened evolution of the OpenClaw personal assistant, it gives every employee a personal AI assistant while specialized “Expert Agents” handle domain-specific tasks like finance or legal review. The platform wraps internal systems (ERP, CRM, databases) into reusable “digital employees” and renders AI output into finished artifacts—dashboards, PPTs, Excel files, and reports—rather than stopping at plain text.

The interesting bit

The unusual angle is that Magic treats enterprise AI risk management as a first-class feature, not an afterthought. Agents run inside proprietary sandbox containers with tenant-isolated VPC networking, high-risk actions trigger human approval workflows, and spending is capped per department, user, or individual agent. It is trying to build the IT security and HR bureaucracy that lets you actually deploy autonomous AI without waking up to deleted databases.

Key highlights

• Renders AI output directly into deliverables (PPTs, Excel, dashboards, canvases) via a built-in framework
• Sandboxed agent execution with sidecar network proxies, private endpoints, and cross-tenant data isolation
• Granular budget controls: daily spending limits per department, user, or agent
• Human-in-the-loop gates for destructive operations like sending emails or deleting data
• Compatible with existing Anthropic Skills and OpenClaw Skills ecosystems

Verdict

Worth evaluating if you are trying to move AI agents from personal experiments to company infrastructure—especially if your CFO worries about API bills and your CISO worries about data leakage. Pure hobbyists and teams without compliance headaches will likely find the guardrails overkill.