The One — a Multi-Agent workshop

Design the agents. Watch the work move.

The Zeroth turns definitions, Graphs, handoffs, task boards and runtime observation into one inspectable operating system.

macOS · Windows

The built-in assistant drives a frontend crew — recorded in The One

The arc

Design. Run. Evolve.

Three recordings straight from The One: author a Graph, let the assistant drive a crew, then edit the crew's genes.

graph-blueprint — the Graph editor

Act 01Design

Draw the collaboration as a Graph.

Bind each node to an Agent preset and write the edges as real handoff instructions — when to hand off, what to pass, what to expect back. Validate, save, and the crew is defined.

  • Nodes bind presets
  • Edges carry handoff instructions
  • Validated, saved, reusable
Graph editor docs
assistant-run — a real delivery

Act 02Run

Hand the work to the assistant and its crew.

Ask in plain language. The assistant stands up a frontend_design_studio sub-Graph and acts as its human — the work chain, the Agent roster and every tool call stay visible until a real website ships.

  • The assistant routes
  • Sub-Graph crews do the work
  • Observable, interruptible
Run docs
a-genetic-loop — the gene-editing loop

Act 03Evolve

Watch it perform, then edit the genes.

The same assistant writes lessons back into the definitions: atlas edits the template's structure, prompts and skills — so the next instantiation of the crew is born better.

  • Templates are genes
  • atlas performs the edits
  • Versioned, reversible
A-Genetic Engineering

Surfaces

Every layer is an inspectable surface.

Engineering ideas

Two engineering ideas.

Prolifera Engineering grows a running system at runtime; A-Genetic Engineering writes a working crew down as a reusable gene.

Prolifera Engineering

The assistant grows the system at runtime.

Prolifera Engineering is how The One builds a multi-agent system as it works — a human-facing chat agent that clones itself, creates agents from presets, stands up whole graphs, and tears them down, modeled on cell proliferation.

Replication

Instantiate a graph, or clone an agent — config, context, and tools — into a fresh identity. The copy is born disconnected and inactive.

Division

One agent becomes several to parallelize bounded work. A long context does not divide an agent; it compresses its memory in place instead.

Differentiation

A created agent is a specialist because its prompt, profiles, skills, and tools are set at creation from its preset.

Apoptosis

A finished agent reports up and is put to sleep; its resources are reclaimed only when a parent explicitly tears it down.

Honest by design: the framework only ever does two things on its own — compress an agent's memory, and sleep a finished one. Proliferation is the agent's own decision and is not globally capped, so it can run away; by default the authority sits in one agent, and the whole system stays observable and can be torn down — not prevented.

Read the docs

A-Genetic Engineering

The crew's gene is a Graph template.

A working crew is written down in full — every agent's prompt, skills, tools, and the edges between them — as a reusable Graph template: the crew's gene. The built-in assistant designs, instantiates, and edits these genes conversationally.

Read the docs

Models & providers

Use the official model service, or bring your own providers.

Official model service

Sign in and run — Starter, Pro and Max plans include managed model capacity, with no API keys to configure and quota packs when you need more.

See pricing

Your own keys & OAuth

Connect providers you already have — OpenAI or Codex OAuth, Anthropic or Claude Code OAuth, Gemini or Antigravity OAuth, and any OpenAI-compatible endpoint.

OpenAICodexAnthropicClaude CodeGeminiAntigravityOpenAI-compatible
Read the docs

Keep exploring The Zeroth.

Read the docs, or browse Graphs, agents and skills the community has shared as Definition Spaces.