Introduction

Zero Operators is an autonomous AI research and engineering team. You write a plan; agents build it; the oracle verifies the work.

What Zero Operators is

Zero Operators (ZO) is not a coding assistant. It’s a digital research-and-engineering team that happens to express itself in code, mathematical models, reports, and data artifacts.

The human is the research director. The plan is the only communication medium. Agents do the rest.

The plan

A single Markdown file describes objectives, success metrics, constraints, milestones — the total brief.

The oracle

Every project defines a hard, verifiable success metric. No deliverable is complete until the oracle confirms it.

The team

A team of 20 specialised agents — orchestrator, data, model, oracle, XAI, and more — coordinates over a contract-first protocol.

The memory

STATE.md, DECISION_LOG.md, PRIORS.md, and a semantic index give every session continuity with the last.

How a project flows

1. Plan approval

   You write (or have an agent draft) a plan.md. Review and approve. 30–60 minutes.

2. Autonomous Phase 1–2

   Agents execute data review and feature engineering. No human involvement.

3. Gate 2 checkpoint

   You review the feature list with a domain expert. 15–30 minutes.

4. Autonomous Phase 3–4

   Agents design, train, and iterate on models autonomously, with the experiment loop deciding when to stop.

5. Gate 4 checkpoint

   You review the full analysis package: performance, explainability, ablations, statistical significance. 30–60 minutes.

6. Delivery

   Agents package everything into a clean delivery repo with zero ZO infrastructure leakage.

The human is involved at plan approval and two blocking gates. Everything else is autonomous. You can edit plan.md at any time to change direction; agents detect the delta and re-plan.

The thesis

ZO combines four sources that haven’t previously been combined:

Karpathy's hard oracle discipline

The plan is the human’s sole lever of control. The system enforces hard oracle discipline: every claim is verified against a concrete, measurable criterion. Autonomy scales through rigorous specification, not natural-language ambiguity.

Tiered model routing + semantic memory (Ruflo)

Three-tier routing assigns Opus, Sonnet, or Haiku based on task complexity. Session hooks maintain agent memory across resumable work. A semantic index over past decisions lets the system learn from its own mistakes.

Contract-first agent teams (coleam00)

All interfaces, inputs, outputs, and success criteria are defined before agents are spawned. Rich spawn prompts encode domain knowledge, precedent, and constraints. Eliminates mid-execution clarifications.

Claude Code native capabilities

Peer-to-peer agent communication via session context. Built-in agent teams. Markdown-based skill and agent definitions. JSONL audit logging.

Where to next

Quickstart

Five minutes from zero to your first ZO run on a real dataset.

Core concepts

The plan, the oracle, agents, phases, gates, memory — start here for the mental model.

CLI reference

Every zo command, every flag, examples for each.

GitHub

Source code, issues, contributing guide.