Distributed AI Agent Framework

A colony of agentsthat thinks.

Antmachine is a distributed mesh of AI agents built on the OODA loop — structured, artifact-driven, and designed to coordinate across instances in real time.

Not a chatbot. A colony.

Get Early Access See What We've Built

Distributed Mesh

Agents sync state across instances in real time via CRDT — Loro under the hood. Multiple nodes, one coherent colony. No central coordinator.

loro-crdt · WebSocket

OODA Loop Core

Every agent runs the same loop: Observe, Orient, Decide, Act. No chaotic jumping to solutions. Artifact-driven — every skill wakes up with only the context it needs.

Claude Code Plugin

Tiered Intelligence

Ant-High (Opus) for deep reasoning. Ant-Med (Sonnet) for execution. Ant-Low (Haiku) for fast tasks. The right model for every job — no wasted tokens.

Opus · Sonnet · Haiku

Agent Architecture

Three tiers. One colony.

Each ant knows what it's good at. The mesh routes work to the right node automatically.

High Priority

Ant-High

model: claude-opus-4

Deep analysis. Novel situations. High-stakes decisions. Security review. Multi-step reasoning. Use sparingly — every call is expensive and should justify the depth.

ReadGrepWriteBashTaskWebSearch

Standard

Ant-Med

model: claude-sonnet-4

The workhorse. Multi-factor analysis, planning, tradeoffs, code understanding. Handles most tasks. Escalates to Ant-High when confidence is low.

ReadGrepGlobWriteBashTask

Fast

Ant-Low

model: claude-haiku-4

Pattern matching, quick checks, speed. Reads and writes. Doesn't attempt complex analysis — escalates when confidence drops below 0.6.

ReadGrepGlobWrite

Observe

Evidence gathering. Each observer forks with a clean context — no conversation history, just the artifacts it needs. Writes an observation report.

→

Orient

Reads observations and history. Detects patterns from prior sessions. Routes to the right next action — decide, act fast, or loop back for more data.

→

Decide

Strategic planning for novel problems. 2–4 options with tradeoffs, not one answer. Evidence-based. Only runs when Orient can't pattern-match a known path.

→

Act

Implementation. Writes artifacts, runs code, updates state. Logs everything to history.jsonl — every action feeds the next Orient cycle.

Built while building Antmachine

Experiments

Real products shipped to prove the framework works. Every one built by an AI agent using the OODA loop.

↗

GymMap

Hoodmaps for gyms. Crowdsourced vibe tags on a live map.

Live

↗

Property.bot