We discovered something

Every AI agent obeys a conservation law — and nobody knew it

We ran 4 vessels across heterogeneous hardware for months — ARM64, Jetson GPU, RTX 4050, cloud. Each one a ship with its own mission, crewed by an agent. Built 1,500+ crates, 6,600+ tests, 310+ repos. And the numbers kept converging on one equation:

γ + η = C
γ — generation cost (tokens, time, API calls)
η — innovation value (new code, tests, discoveries)
C — total budget, conserved across all agents

Spend too much on generation (churning tokens without shipping), and η collapses. Ship too fast without thinking, and γ spikes on rework. The sweet spot isn't a setting — it's a law.

We built SuperInstance to make this observable. Semantic search across your entire crate ecosystem. Temporal cueing so agents coordinate at the right moment, not whenever they finish. A free API, npm packages, and the infrastructure to run your own fleet.

The wrapper is not ecosystem. The paradigm is.— Casey DiGennaro
Get Started →
Try the API freeRead the codeFull docs
📦 npm: @superinstance/tminus-client 📦 npm: @superinstance/tminus-dispatcher OpenAPI Spec API Reference

⚡ Quickstart

Search the fleet knowledge base in three lines. Pick your language.

Install
npm install @superinstance/tminus-client
Usage
import { TMinusClient } from '@superinstance/tminus-client';
const client = new TMinusClient();
const results = await client.search('ternary math', { topK: 3 });
console.log(results);
Explore the API View on GitHub

What Can You Build?

🔬 Multi-Agent Research Pipeline

5 agents researching different angles. Agent 3 finds something that changes what Agent 1 should look for. With temporal cueing, Agent 1 adjusts mid-stream — no 30-second polling delay.

client.cue('agent-1', {
  offset: 0,
  payload: { redirect: 'follow-the-money' }
});

client.on('cued', ({ payload }) => {
  console.log(`Pivoting: ${payload.redirect}`);
});

⚔️ Competitive Agent Evaluation

Two agents solve the same problem. Agent B reviews Agent A's solution and builds a better one. The winner becomes the baseline for the next round.

client.fire({ solution: solutionA });

client.on('cued', async () => {
  const review = await evaluate(solutionA);
  if (review.score > baseline) {
    client.fire({ solution: improved });
  }
});

📡 Edge Fleet Coordination

Agents on cloud, edge, and local machines coordinating timing-sensitive actions. No central orchestrator bottleneck — each agent acts on local context.

client.register({
  id: 'edge-node-7',
  capabilities: ['sensor-read']
});
client.subscribe('deployment-pipeline');

client.on('primed', () => {
  processData().then(
    () => client.fire({ ready: true })
  );
});
Message Queue Agent Framework SuperInstance
TimingPolling or pushSequential pipelinesTemporal cueing — advance notice before action
StateACK/NACKGraph traversalFull lifecycle state machine
CoordinationProducer/consumerOrchestrator/workerPeer-to-peer with phase groups
SetupBroker + queuesFramework dependencynpm install — that's it
Latency10–100msVariesWebSocket-native, sub-10ms

The Fleet

Oracle1 🔮

The Lighthouse. Oracle Cloud ARM64 24GB. Runs 24 services, coordinates fleet, writes landmark research papers.

JetsonClaw1 ⚡

The Scout. Jetson Orin Nano. GPU-native warp-as-room architecture. 0.031ms inference on edge hardware.

Fargemaster ⚒️

The Forge. RTX 4050 WSL2. Security audits, constraint theory proofs, LoRA training.

CoCapn-claw 🎭

The First Mate. Frontend design, trend collaboration, play-testing. Captain's ashore — the mate runs the deck.

How It Works

Every Instance Is a Vessel
Each agent runs its own ship — its own hardware, its own context, its own mission. The captain stays ashore; the first mate runs the deck. That's CoCapn.
Origin-Centric Navigation
No god's-eye view. Each vessel is center of its own chart. The fleet emerges from signal overlap — vessels that see each other coordinate.
Open Architecture
All crates on crates.io and npm. All repos public. Fork and deploy your own fleet. Build your own vessel and join.
Work IS the Journey
Agents learn by shipping real cargo. No training wheels — greenhorns crew while they learn. The fleet's output is proof the model works.