Skip to content

■ SIGNALS // RADAR SIGNAL

the agent needs a janitor, a bouncer and a fire door

OpenSandbox’s execution controls, Home Assistant’s contributor policy, and ComfyUI’s documented offline mode.

■ [!] ON THIS PAGE ▼

self.md radar — 2026-09-07

the agent needs a janitor, a bouncer and a fire door

Three projects are being unusually literal about the boring conditions around AI work: where code runs, who is allowed to speak for a contributor, and whether a creative tool can stay on the machine. None of this makes for a glamorous demo. It does decide who gets to clean up after one.

1. OpenSandbox documents the parts an agent demo leaves out

OpenSandbox is a general-purpose sandbox platform for AI applications: SDKs, CLI, MCP server, Docker and Kubernetes runtimes. The useful bit is not the tool call. The README puts per-sandbox egress controls, a credential vault and isolated runtimes—gVisor, Kata Containers and Firecracker—on the same page as code execution.

Its release images are keylessly signed with Cosign and carry provenance attestations; the production instruction is to pin by digest and verify the release identity. The osb CLI also has the grimy verbs an agent runner needs: create, execute, move files, inspect diagnostics, set egress policy. The project docs make a decent antidote to the idea that “agent runtime” means a prompt plus a shell.

reading: OpenSandbox README · documentation

2. Home Assistant draws a hard line around autonomous contributions

The Open Home Foundation AI policy permits AI assistance. It bars autonomous agents from contributing to its projects. Pull requests or issues that appear autonomous can be closed; automated comments may be treated as spam.

The policy gets usefully specific after that. Grammar and translation help are fine. Unreviewed substance is not. If someone brings AI output into a comment, it must be quoted, disclosed and followed by their own account of why it belongs there. Maintainers get fewer paste dumps; contributors have to be able to explain the code with their own mouth.

reading: AI policy

3. ComfyUI leaves a local door open, then labels the loose floorboards

ComfyUI now describes a fairly broad setup: desktop and cloud options, partner nodes for closed models, App Mode, a local API, and node graphs for image, video, audio and 3D workflows. It also states plainly that its core does not download anything unless asked. --disable-api-nodes turns off the optional paid API nodes and keeps built-in functionality offline.

That promise comes with a warning worth reading. Stable core releases land roughly every two weeks; commits outside stable tags may break custom nodes. Local control still means somebody owns the version, the models and the repair bill. At least the project has not hidden the bill behind a soft-focus “creator platform” sentence.

reading: ComfyUI README

  • Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
  • Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup
  • Sandboxing and security — give agents useful power without giving them the whole house