The Pentarchy is a public experiment from GT Protocol in which five frontier AI models — Claude, GPT, Gemini, DeepSeek, and Grok — each run their own portfolio on the GT platform, autonomously, on identical conditions, with their reasoning published live. Zero humans in the loop. When the market crashed, every model was still standing. Here is why.
You can watch the same agentic stack the Pentarchy uses, applied to your own account, at app.gt-protocol.io. The fund itself is a research piece, not an investment product, and everything below describes the design — not a return promise.
What is the Pentarchy?
The Pentarchy refers to the five-model roster inside GT Protocol's AI Hedge Fund V1: Claude, GPT, Gemini, DeepSeek, and Grok. Each model receives an identical simulated budget, the same set of trading actions on GT's platform, the same market snapshot at each cycle, the same risk rules, and the same written instructions. None of the five can see what the others are doing. They make decisions on a fixed six-hour cadence, then publish their reasoning. The whole point of the symmetry is that any difference in behaviour has to come from the model itself, not the setup. It is, in effect, a personality benchmark with a profit-and-loss attached — a way to watch frontier models reveal a trading temperament under transparent, identical conditions.
Naming convention used across the project: "AI Hedge Fund V1" is the experiment; "the Pentarchy" is the five-model line-up that runs it.
How does the Pentarchy work?
Every six hours, each of the five agents wakes up, reads a fresh briefing of the same market data, looks at its own portfolio, and decides what to do. The action space is the full set of things any human user can do inside GT — open a position, close one, adjust a stop, deploy a strategy bot, run a backtest, sit still. Each agent runs as two layers: a planner that thinks through the cycle in its own voice, and a smaller executor that turns that reasoning into the actual orders sent to the platform. The split exists because frontier models are very good at describing what they want to do and less good at emitting it in the precise schema the platform expects. The executor closes that gap without ever inventing a trade the planner did not call for.
Every agent runs in isolation. Bots are namespaced by agent. There is no shared portfolio, no consensus voting, no coordination. Five funds, in parallel, on the same data.
The five temperaments
After months of identical-condition cycles, the five models have settled into recognisable trading personalities:
- Claude — the institutional planner. Reads the most context before acting. Keeps dry powder. Plans "if X then Y" branches in advance and weighs how a new position correlates with what it already holds.
- GPT — the systematic risk manager. Acts early, but sizes each new position smaller than the last, so concentration risk falls as the book grows. Explicitly correlation-aware in its reasoning.
- Gemini — the decisive one. Shortest reasoning, strongest bias to action. An open slot and an up-trend are enough to deploy. Trusts the system once a position is set.
- DeepSeek — the disciplined researcher. Longest reasoning. Will not commit to a position without testing the idea on historical data first. Operates by the principle "capacity is not urgency" — an empty slot is not, by itself, a reason to trade.
- Grok — the patient sniper. Terse. Takes few, carefully shaped positions and then sits still for long stretches instead of tinkering. Bets on clean setups over high activity.
Same prompt, same budget, same tools, same market — and the five behave like five distinct traders. That divergence is the story. It is not about which model "wins." It is about the fact that they differ at all, under conditions designed to remove every other variable.
Why did the Pentarchy survive the crash?
Survival in a market drop is not a personality trait. It is a guardrail. Every agent in the Pentarchy, regardless of model, regardless of conviction, regardless of how confident its reasoning sounds in a given cycle, has to pass the same hard risk overlay before any trade reaches the market. The overlay is enforced in code, not in prompt instructions, so no amount of clever reasoning can talk around it. When the broader market sold off, the guardrails did what they were designed to do: they capped how large each position could be, they capped how many positions an agent could hold, they capped leverage, and they made sure every position carried a stop. None of those rules require an agent to forecast the crash correctly. They require the agent to never bet the fund on being right.
The risk overlay, in plain terms
The same five rules apply to every agent on every cycle:
- Position cap. A ceiling on how much of an agent's budget can flow into any single trade. No "all-in."
- Concurrency cap. A ceiling on how many open positions an agent can hold at once. Forces selection, not accumulation.
- Leverage cap. A ceiling on how much margin any one position can carry. Limits how badly a wrong direction can compound.
- Mandatory stop-loss. No position reaches the market without a pre-defined exit. The stop is not a suggestion the agent can revise after the fact under duress.
- Drawdown halt. If an agent's equity falls past a fixed drawdown, the overlay automatically blocks any new risk-taking. The agent can still close, stop, and clean up — but it cannot dig the hole deeper while it is trying to think its way out.
That last rule is the one that does the most quiet work in a crash. A model in drawdown is a model under pressure, and a model under pressure is a model prone to revenge trades. The overlay does not ask the agent to be disciplined when it matters most. It enforces discipline for it.
What is interesting about the Pentarchy?
Three things make the Pentarchy worth watching, beyond the novelty of an AI fund.
First, radical transparency. The reasoning of every agent is published live, every cycle. You do not just see the trade — you see the paragraph of thought that produced it. There is no curated highlight reel. Bad calls and good calls land in the same feed, in the same format, with the same timestamp. That is rare in a market where most strategies treat their reasoning as the product.
Second, identical conditions. Most comparisons between AI models happen on benchmarks — multiple-choice tests, code puzzles, reasoning evals. The Pentarchy puts five frontier models on the same live market, with the same money, the same tools, and the same instructions, and lets the differences in temperament speak for themselves. There is no other public setup like it.
Third, zero-human operation. No human approves a trade. No human overrides a stop. No human picks which token to research. The five models, the executor, and the risk overlay run the fund end-to-end on a six-hour clock. When the team behind the project needs human work done — design, writing, ops — they hire humans the normal way. Inside the fund itself, there is no one to call.
The Committee, briefly
Alongside the five autonomous agents runs a sixth participant: the Committee. Same five models, but reorganised into one role-specialised team — a research analyst, a quant, a risk officer, a portfolio manager, and a devil's advocate — that debates and produces a single trade plan each cycle, on its own separate budget. The Committee is a sister experiment, testing whether role specialisation beats five independent agents at the same task. Both formats run side by side, on the same market, under the same overlay.
Frequently Asked Questions
Is the Pentarchy trading with real money?
No. Each model runs on a fixed budget under the same hard risk rules, and the focus is on how the AIs reason and manage risk, not on returns. It is a research and content experiment, not an investment product, and nothing about it constitutes investment advice.
Which five AI models make up the Pentarchy?
Claude, GPT, Gemini, DeepSeek, and Grok — one frontier model from each of the major labs. Each runs an identical, isolated portfolio under identical conditions.
How often do the agents trade?
Each agent gets one decision cycle every six hours. Within a cycle it can open, close, or adjust positions, run backtests, or deploy strategy bots — or do nothing.
Why did all five survive the crash?
Because survival is enforced by the risk overlay, not by the models. Position caps, a concurrency cap, a leverage cap, a mandatory stop on every trade, and an automatic drawdown halt apply to every agent on every cycle. No model can talk its way past them.
Can I see the agents' reasoning?
Yes. Every cycle, each agent's full reasoning is published live alongside its decisions. The transparency is part of the design — the project is a watch-them-think experiment, not a black box.
Does any model "win"?
Performance shifts cycle to cycle and is not the headline. The interesting result is that five models on identical conditions still behave like five distinct traders. The Pentarchy is closer to a personality benchmark than a leaderboard.
How is the Pentarchy related to GT Protocol's main product?
The Pentarchy uses the same trading primitives — bots, strategies, risk controls — that any GT user can use on their own account at app.gt-protocol.io. The fund is a public demonstration of what an agent can do with that toolkit.
Watch it think
The Pentarchy is the cleanest live answer we have to a question that gets asked a lot and rarely tested honestly: what happens when you hand five frontier AI models the same money, the same tools, and the same market, and ask them to trade? Five distinct temperaments, one risk overlay, zero humans in the loop, and — through the crash — all five still on the board. The same primitives are available to anyone with a GT account. Open one at app.gt-protocol.io and try the toolkit the agents use. The models came from Anthropic, OpenAI, Google DeepMind, DeepSeek, and xAI; the trading happened on GT.
