All work

Internal R&D · the build engine

Agent-OS

A deterministic, self-healing multi-agent system that plans, builds, and verifies production software end-to-end — the engine behind our own products.

AI Development·Cloud Architecture
12
agents orchestrated
755
tasks defined
98.7%
verified autonomously
$0
infra — Cloudflare free tier

The hard part of agentic systems isn't getting an LLM to do a task once. It's getting a fleet of them to do the right tasks, in the right order, reliably — and to know when they've actually succeeded. Agent-OS is the deterministic system that does it, and the engine behind our own builds.

How it works

A poll → execute → validate → unblock loop coordinates a fleet of agents across a dependency graph of tasks.

  • A zero-trust validator sits between every step — no agent is trusted to grade its own work.
  • Durable Objects hold stateful coordination; D1 is the task ledger.
  • Workers AI + Vectorize provide grounded retrieval so decisions are anchored, not hallucinated.
  • Real-time WebSocket telemetry makes every decision the swarm makes observable.

The result

755 tasks defined, 745 verified autonomously — 98.7% — at zero infrastructure cost on Cloudflare's free tier.

  • This is the system that ships our products: Kointel, KTHULHU, and Art of Zeta all draw on it.
  • Deterministic where it counts — an engineering problem solved with engineering, not a bigger prompt.

Built with

Durable ObjectsD1Workers AIVectorizeWebSockets

Want one built like this?

The engineering behind Agent-OS is available for your system.