Overview
TensorQ is an autonomous AI agent that trades Bittensor subnet alpha tokens. It monitors on-chain data across every subnet, identifies mispriced alpha tokens, executes trades, and learns from the results.
In Bittensor's dTao system, each subnet has its own token called "alpha." When you stake TAO to a subnet, you receive alpha tokens. The alpha price moves based on supply and demand — if more people stake to a high-performing subnet, the alpha appreciates. If they leave, it drops.
This creates a market with real fundamentals. Subnets that earn more emissions, attract more miners, and deliver better results should see their alpha appreciate. But the market doesn't always price things correctly. Subnets get overlooked. Narratives lag reality. Registration spikes precede price moves by hours or days.
TensorQ reads this data continuously and acts on it. Not on hunches — on measurable, on-chain signals that have historically correlated with alpha price movements.
What makes it different
- Fully transparent. Every trade the agent makes is visible. The reasoning behind each position is published in real-time. You can judge the agent's logic before committing anything.
- On-chain data, not speculation. The agent doesn't follow social media sentiment or influencer calls. It reads the blockchain — emissions, stake flows, registrations, validator consensus — and makes decisions based on what the network is actually doing.
- Self-improving. After every trade cycle, the agent reviews what worked, what didn't, and adjusts its strategy weights. A signal that led to five consecutive losses gets downweighted. A pattern that predicted three winners in a row gets promoted.
How the Agent Thinks
The agent operates on a continuous loop with six phases. Each cycle runs on a configurable interval — typically every few minutes during active market conditions, less frequently when the network is quiet.
1. Scan
The agent connects to the Bittensor Finney chain via WebSocket RPC and pulls the latest state for every tracked subnet. This includes emission allocations, total stake per subnet, net stake changes over the last epoch, new miner registrations, validator consensus scores, and alpha token prices from DEX data.
The scan produces a structured snapshot — a point-in-time view of the entire network that the agent can reason about.
2. Analyze
The snapshot goes to Claude, along with the agent's current positions and recent trade history. The AI evaluates each tracked subnet against its current portfolio: Is this alpha undervalued relative to its emission share? Is stake flowing in faster than the price has moved? Has the registration rate spiked — a pattern that historically precedes price increases?
The analysis produces a ranked list of opportunities with confidence scores.
3. Decide
Based on the analysis, the agent decides what to do. This might be opening a new position (buying alpha on an undervalued subnet), closing an existing one (taking profit or cutting a loss), or doing nothing (no high-conviction opportunities). The agent also considers its current exposure — it won't overload into a single subnet regardless of conviction.
4. Execute
If the decision involves a trade, the agent submits the transaction directly to the Bittensor chain via polkadot.js. For buys, it stakes TAO to the target subnet to receive alpha tokens. For sells, it unstakes alpha to receive TAO back. Every transaction is on-chain and verifiable.
5. Journal
After every decision — including the decision to hold — the agent writes a journal entry. This records what it saw, what it considered, what it decided, why, its confidence level, and what it expects to happen. This journal is the foundation of the transparency system.
6. Learn
Periodically, the agent reviews its closed trades and compares outcomes to expectations. If it expected a subnet's alpha to appreciate within 48 hours based on a stake velocity signal, and instead the price dropped, it records that outcome. Over time, this feedback adjusts the weights the agent assigns to different signals.
Signals & Data
The agent watches a specific set of on-chain metrics. Each signal has been chosen because it has a logical relationship to alpha token value — not because it backtested well on random data.
| Signal | What it measures | Why it matters |
|---|---|---|
| Emission Share | Percentage of total TAO emissions going to a subnet | More emissions = more fundamental value flowing to the subnet. If alpha price hasn't caught up, the token may be undervalued. |
| Stake Velocity | Net TAO staked/unstaked over a rolling window | Leading indicator. Smart money moves before price. Consistent inflows often precede alpha appreciation. |
| Registration Rate | New miners joining the subnet per epoch | Growing miner interest signals subnet demand. Registration spikes have historically correlated with price moves within 24-72 hours. |
| Validator Consensus | Agreement among validators on miner scoring | High consensus = healthy, stable subnet. Low consensus = potential instability or gaming. The agent avoids low-consensus subnets. |
| Alpha Price | Current TAO/alpha exchange rate on the dTao market | Compared against fundamental value (emission backing) to identify over/underpricing. Price alone means little without context. |
| Volume & Liquidity | Recent trading volume and available liquidity depth | The agent must be able to enter and exit without excessive slippage. Low liquidity subnets are avoided regardless of opportunity. |
| Subnet Age | How long the subnet has been active | Newer subnets are higher risk/higher reward. The agent adjusts position sizing based on maturity. |
| Cross-Subnet Correlation | How alpha tokens move relative to each other | Diversification signal. The agent avoids concentrating in subnets that tend to move together. |
Risk Management
Making good trades is only half the problem. Surviving bad ones is the other half. The agent has hard-coded risk rules that cannot be overridden by its own reasoning.
Position sizing
The agent sizes positions based on confidence. A high-conviction trade (75%+ confidence) gets a larger allocation than a speculative one (55-65%). But no single position can exceed 20% of the total portfolio, regardless of confidence.
Stop-losses
Every position has a stop-loss. If an alpha token drops 15% from entry, the position is closed automatically. The agent doesn't "hope" or "hold through the dip." The stop is hard — it fires regardless of the agent's current analysis.
Cash reserve
At least 20% of the portfolio stays in TAO at all times. This ensures the agent can always act on new opportunities and isn't forced to sell existing positions at bad prices to fund new ones.
Diversification
The agent tracks cross-subnet correlation and avoids concentrating in subnets that tend to move together. If SN8 and SN19 are highly correlated, taking large positions in both effectively doubles the risk. The agent treats correlated subnets as partially overlapping exposure.
Drawdown protection
If the portfolio drops 10% from its peak in a rolling 7-day window, the agent reduces all position sizes by 50% and enters a cautious mode. It stays cautious until it recovers to within 5% of the peak. This prevents catastrophic losses during broad market downturns.
Performance
All performance numbers shown on the homepage and in the app are calculated from actual on-chain trades. Nothing is backtested or hypothetical.
How PnL is calculated
For each position, PnL is straightforward: the TAO received when selling alpha minus the TAO spent when buying it. For open positions, unrealized PnL uses the current alpha/TAO rate. Total portfolio PnL is the sum of all realized and unrealized position PnL.
Equity curve
The equity curve shows the total portfolio value (in TAO) over time. It's snapshotted every hour. The curve includes both the TAO held in reserve and the current market value of all alpha positions.
Win rate vs profit factor
Win rate alone is misleading. An agent that wins 90% of trades but loses 10x on each loss is unprofitable. The profit factor — total profits divided by total losses — is a better measure. A profit factor above 1.5 means the agent is meaningfully profitable. Above 2.0 is excellent.
$TENSORQ Token
$TENSORQ is an ERC-20 token on Base that funds the trading agent's operations and gates access to the full dashboard.
How fees work
When $TENSORQ is traded on DEXes, a small percentage of the trading fee is collected by the protocol. These fees are periodically converted to TAO and deposited into the agent's trading wallet. This creates a direct link between token activity and the agent's trading capital.
What holders get
- Full dashboard access: Live positions, AI reasoning for every trade, strategy weights, complete trade journal.
- Strategy insights: What signals the agent is currently prioritizing, what it's learned from recent trades, what it's watching next.
- Subnet scanner: Real-time view of all tracked subnets with the same metrics the agent uses.
Token details
- Chain: Base (ERC-20)
- Launch: Via launchpad (contracts not in this repo)
- Access threshold: Hold a minimum amount to unlock full app
Transparency
Most trading bots are black boxes. You send money, hope for the best, and maybe see a PnL number. TensorQ is the opposite.
Every trade has a reason
When the agent opens or closes a position, it publishes the full reasoning: what signals triggered the decision, what it expects to happen, its confidence level, and how the trade fits into its overall strategy. This isn't a vague summary — it's the actual analysis the agent performed.
On-chain verification
Every trade is a transaction on the Bittensor chain. You can verify the agent's wallet address, see its stake positions across subnets, and confirm that the trades match what the dashboard shows. The agent's wallet is public.
Trade journal
The full trade journal is available in the app. For every position, you can see: the entry reasoning, any updates during the hold period, the exit reasoning (or current status for open positions), and the actual outcome vs what the agent expected. Over time, this builds a complete record of the agent's decision-making quality.
Why transparency matters
Trust in autonomous agents is earned, not assumed. By making everything visible, TensorQ lets you evaluate the agent's thinking before you decide whether to support it. If the reasoning is sloppy, you'll see it. If the strategy is sound, you'll see that too.