Sandboxes for Coding Agents | Forward Planning #12

Forward Planning is a live stream about coding with AI in practice: workflows, tools, and how engineering is changing as AI becomes part of the development loop. In Episode 12, we look at sandboxes: why they matter for AI agents, and what different sandboxes can look like depending on what you're trying to do. When an agent is writing code, running tests, or submitting a pull request, it needs somewhere safe to do that work. But "sandbox" means very different things in practice: a container, a microVM, a full virtual machine, an isolated filesystem. The right choice depends on how much isolation you need, how much you trust the agent, and whether the laptop it's running on is even on. We'll walk through the tradeoffs between local and remote sandboxes, talk about the pets-vs-cattle mindset for agent environments, and dig into some concrete scenarios. Topics we'll cover: - Why agents need sandboxes and what can go wrong without them - Local sandboxes: containers, microVMs, VMs, and filesystem isolation (AgentFS) - Remote sandboxes: what they unlock when your laptop is closed - Pets vs cattle: how to think about agent environments As always, join us live to ask questions in chat and share how you're sandboxing your own agent workflows.