Preview on SAM

Sandboxed agents

Mon, 01 Jan 0001 00:00:00 +0000

Preview

The sandbox works end to end and is exercised in CI, but sam-box flags, the bundle format and nano-init’s options may change between releases.

An agent is a harness, a model and some tools. To deploy one you normally ship all three, plus a key for the model, plus network access that is wide enough for whatever the tools turn out to need. On SAM you ship only the harness. The mesh provides the model and the tools by name, and the network the agent gets is a list of names that is checked on every connection.

Agent architecture

Mon, 01 Jan 0001 00:00:00 +0000

Preview

This is the design behind sandboxed agents. The decisions here are settled; the interfaces that implement them are still being shaped, and the last section lists what is not built yet.

This page explains how an unmodified agent, running inside a sandbox with no network of its own, is connected to the mesh. It covers the layers, the reasons for each layer, the contracts between them, and the limits of what the design guarantees. The scale report measures what the result costs.

SAM Connect (Android and iOS)

Mon, 01 Jan 0001 00:00:00 +0000

Preview

The app is functional and tested in CI on an Android emulator. Its screens, the on-device tools it exposes and the FFI surface may change.

SAM Connect is a Flutter app that runs sam-node on a phone. The Go node is compiled into a shared library and driven over Dart FFI. The app is the UI around it, plus a small MCP server that exposes the phone’s sensors to the mesh. A phone enrolled this way is a node like any other. It has a key, a credential and a peer ID, and it appears in discovery with its phone-sensors service.

Scale report: what an agent costs

Mon, 01 Jan 0001 00:00:00 +0000

Preview

These measurements are of the sandbox datapath described in Sandboxed agents. They were recorded on 2026-08-21 at commit e5b2966, when the boundary spoke SOCKS5. The boundary has since moved to HTTP CONNECT and has not been measured again. The shape of the results is what matters. Single figures are indicative.

The design gives every agent its own boundary, which resolves names and enforces policy on each flow. That sounds expensive. This page measures it. The method is written down so that the result can be challenged, and a script is provided so that it can be run again. There are two experiments: a sweep of boundaries attached to one node, and a thousand agents in microVMs on one host.

Scale experiment: reproducing the thousand-agent run

Mon, 01 Jan 0001 00:00:00 +0000

Preview

The scripts under tests/scale/ and scripts/ are research tooling for the scale report. They are kept working but are not a supported product surface.

The report says what was measured and what it means. This page says how to run it again. The setup is one host, one sam-node, and N agents. Each agent is a Firecracker microVM with no network device, and its only way out is a vsock to its own sam-box. A thousand agents have been run on one n2-standard-64.