A microkernel for your agent swarm

What it does

AgentOS is a C-based systems stack that treats multi-agent orchestration as an operating-system problem. It provides a microkernel (atoms, about 6 KLOC), ten-plus daemon services, a JSON-RPC gateway, and language SDKs so teams can coordinate agent swarms via syscalls rather than framework callbacks. The full repository clocks in around 478 KLOC and includes everything from sandboxing to a desktop application layer called openlab.

The interesting bit

Rather than tacking security and memory onto a Python script, AgentOS builds them into the lower layers: sandbox isolation, RBAC with a YAML rule engine, and a four-tier memory model (raw → feature → structure → pattern) that aims to stretch context-window limits. The open-source build only includes the first two memory tiers; the full L1–L4 stack requires the proprietary MemoryRovol component.

Key highlights

• Microkernel design (~6 KLOC kernel, ~478 KLOC total) with strict layering from syscalls up to HTTP/WebSocket gateways
• Four-layer “inherent security”: sandboxed workspaces, RBAC/YAML permission arbitration, input sanitization, and tamper-proof audit trails
• Tiered memory model; OSS users get L1–L2, while L3–L4 are reserved for the Pro tier
• Native SDKs for C/C++, Python, Go, Rust, and TypeScript
• Ships as v0.1.0 preview builds for both Docker and desktop

Caveats

• The README makes grand historical claims (“seminal fourth OS philosophy”) and gives contradictory token-efficiency figures (“500 % savings” in one section, “80 %” in another) without benchmarks or methodology
• Most detailed documentation links are dead placeholders, so the 24 core principles and five-dimensional orthogonal system are described but not demonstrated
• It is an early v0.1.0 preview; the distance between the architectural manifesto and hardened code is left to the reader’s imagination

Verdict

Systems programmers building low-level agent infrastructure may want to track the repo for its kernel-style isolation. Developers looking for a quick, drop-in agent framework will likely find it over-engineered and under-documented for now.