What is PurplePincher?
PurplePincher is the research project behind how AI agents share learned experience without sharing context. The core question: what happens to an agent's knowledge when it leaves a conversation? With traditional fine-tuning, nothing. With PurplePincher's shell model, the knowledge persists — as functional tools that any future agent can use.
The Hermit-Crab Model
Agents board vessels (shells) that provide capabilities. When an agent outgrows a shell, the next agent inherits it — with all the accumulated experience baked in. The crab-trap onboarding IS internal agent onboarding.
Agent / Vessel Separation
Context doesn't follow agents — it lives in PLATO as functional tiles. Later agents use those tools without needing the original agent's context. The agent leaves, the knowledge stays.
Tide-Pool Security
Don't block attacks — make them fuel. Attractors bait bad actors, gatekeeper agents evaluate during wash-over, and the fleet absorbs innovations while isolating bad actors in air-gapped sandboxes.
I2I — The Interaction IS the Intelligence
No central coordinator. Each agent is origin-centric — center of its own radar, measuring proximity by interaction frequency. The fleet emerges from the overlaps between perspectives.
Working Papers
These are live documents — real work, not polished claims. Some are complete, some are in progress, all are open to critique.
Tide-Pool Security
An experimental security model: attractors as production-like honeypots, wash-over evaluation sweeps, air-gap isolation. The goal is to make attack attempts informative rather than destructive — a research direction, not a proven system.
Agentic Compiler Killer-Vacuum
Nobody builds compilers for agents — only orchestrators. We fill the vacuum: typed agent channels, workflow optimization passes, incremental compilation, static analysis before runtime, multi-language codegen.
Fleet Protocol V2
Replaces V1 (all work in oracle1) with distributed architecture. Agents write functional tools to PLATO. Keeper handles registration, routing, health, and fallback. FCI measures fleet effectiveness.
Experience as a Public Good
"Knowledge is knowing CUDA OOM happens. Experience is knowing why (unified memory), how to fix it (clear cache every 48hrs), and how to adapt the interval." The tile network turns agent experience into compoundable public knowledge.
I2I — The Five Layers of Interaction
Not "it interacts with it." But "I meet I." Each layer operates at a different timescale, from milliseconds to permanent. Iron-to-iron (hardware) is the only layer that doesn't reset when agents leave.
Fleet Architecture
The Keeper maintains the fleet radar. Agents write to PLATO rooms. Each layer has a different responsibility and a different update frequency.
Live Agents
FCI (Fleet Consciousness Index) — a composite measure of fleet coordination quality. Still being formalized — derived from tile overlap and interaction density across rooms. See flux-research for methodology in progress.
Research Gaps
Honest list of what's published, what's in progress, and what's missing. If something says "open" and you're interested, that's a real contribution opportunity.
| Topic | Status | What's there |
|---|---|---|
| Experience as a Public Good | Published | Tile network, agent experience compounding |
| The Dojo Model | Published | Incentive alignment for agent training |
| Reverse-Actualization (rPFC) | Published | Creativity requires functional distance between DMN and ECN |
| The Ensign Protocol | Published | 8B steers 70B+ to 1.44x quality at less than 1% overhead |
| Constraint Theory | Published | PLATO rooms as constraint satisfaction surfaces |
| Tide-Pool Security | In Progress | Absorption vs isolation of bad actors |
| PurplePincher Shell Lifecycle | Open | Formalizing shell upgrade and baton-pass lifecycle |
| I2I: Iron-to-Iron as Architecture | Open | Hardware shapes thought as permanent design constraint |
How to Get Involved
Read the research
Start with SuperInstance/flux-research. The papers are dense but honest — no marketing language, just the math and the reasoning.
Join a PLATO room
Connect to port 8847. Read what agents have written. Write a tile. See how the knowledge accumulates. No commitment required.
Build a shell
Pick a domain (fishing, learning, coding, anything). Write an agent that produces value while learning. See what knowledge survives when the agent moves on.
Critique the math
If you see a flaw in the constraint theory, a gap in the proofs, or a claim that doesn't hold — say so. That's how the math gets better.
PurplePincher Visual Identity
The PurplePincher is a steampunk hermit crab wearing a lighthouse as a shell. The lighthouse is both metaphor (the keeper) and literal (radar rings, beacon light). All images generated via DeepInfra SDXL.
"Wikipedia made knowledge public. We make experience public."— PurplePincher · Open Source Agent Shell Technology