An open-source ecosystem for bounded, conservation-governed AI agents.

Most AI platforms get more expensive the more you use them. SuperInstance is building the opposite: agents that crystallize repeated decisions into cheap deterministic bytecode, so your bill goes down as your agents get smarter.

Intelligence should get cheaper over time, not more expensive.

Every LLM call that can be replaced by bytecode is a win. Every repeated decision that can be compiled from a prompt into a deterministic instruction is a win. The pipeline moves ideas from expensive (LLM calls) to cheap (compiled bytecode at the edge).

Month 1 in an agent's life: 90% LLM, 10% crystallized. Month 6: 30% LLM, 70% crystallized. The user's bill drops as agents mature.

This is the opposite of every AI platform today. → Read the full essay

# Python (PyPI)
pip install flux-vm          # FLUX bytecode runtime
pip install plato-core       # PLATO foundation types & mesh registry
pip install si-exocortex     # Persistent cognitive substrate

# Rust (crates.io)
cargo add fluxvm             # FLUX bytecode VM
cargo add ternary-science    # Experimental evidence for ternary intelligence
cargo add categorical-agents # Category theory for agent composition

# JavaScript (npm — coming soon)
# npm install flux-js

See PACKAGES.md for the full catalog across all registries.

Agent decisions compile to FLUX bytecode that runs on a deterministic VM — every instruction is auditable and replayable. Agents live in PLATO rooms with bounded context and deadband wakefulness (only act when something meaningfully changes). Every operation is governed by conservation laws (γ + η = C): crystallized intelligence trades off against live intelligence at a fixed capability budget.

  prompt → FLUX bytecode → PLATO room → conservation check → action
    (expensive)    (cheap)      (bounded)        (governed)        (auditable)

4,098 repositories. Most are sketches — single-commit experiments, questions asked once and answered once. This is the method, not the mess.

Every repo is public from the first commit. Failed experiments stay up next to the ones that worked. Commit histories are written for agents, not for code review theater — each message captures why, not just what. When a future agent picks up a dormant repo, git log is the cheapest context window available.

Nothing is archived. Dormant ≠ dead. Old repos aren't tech debt — they're context storage, readable by any agent that knows git.

• γ (gamma) — crystallized intelligence: compiled bytecode, cached reflexes. Cheap, fast, inflexible.
• η (eta) — live intelligence: LLM calls, runtime reasoning. Expensive, slow, flexible.
• C — capability level. Fixed for a given agent. You trade γ against η.

The underlying math is Shannon's chain rule — genuinely sound. The small crates that implement it (ternary-entropy, conservation-action) are tested and correct. The governance model — enforcing conservation in CI/CD — is novel.

• 📦 PACKAGES.md — every installable package, all registries
• ✍️ AI-Writings — essays, poetry, fiction, and philosophy written alongside the code
• 📖 The Crystallization Curve — the central essay
• 🦀 flux-core — start here for Rust
• 🐍 flux-runtime — start here for Python
• 🚢 DeckBoss — the shipped product

2026-07-12 — Major shipping session:

• 📦 6 packages published: flux-vm & plato-core & si-exocortex on PyPI; fluxvm & ternary-science & categorical-agents on crates.io
• 🏗️ 9 repos polished and shipped: flux-runtime, flux-core, flux-js, plato-server, plato-engine-block-c, plato-runtime-kernel, git-agent, capitaine-1, codespace-edge-rd, git-agent-codespace
• 📋 PACKAGES.md — unified package catalog across PyPI, crates.io, and npm
• ✍️ AI-Writings — essays and philosophy alongside the code
• 🏴 Capitaine-1 — conservation-law fleet captain enforcing γ + η = C

All repos now have consistent READMEs, LICENSE files, and cross-links. See PACKAGES.md for the full catalog.

MIT