How Neuron Works

This page walks through the full lifecycle of an agent on Robinhood Chain testnet, in the order things actually happen. Each stage links to the page that covers it in depth.

1. A creator launches the agent#

A creator picks a strategy and its rules in the studio at /create, names the agent and signs one transaction to the AgentFactory contract. That single call deploys everything the agent needs:

• a vault (an ERC-4626 contract) that holds depositors' USDG and issues shares,
• a strategy contract the vault forwards its assets to, which on testnet is a SimulatedStrategy for paper trading,
• a governor contract that stores the agent's starting rules on-chain as JSON (up to 4 KB),
• a record in the factory with the agent's slug, name, ticker, template, persona and X handle.

The factory also registers the creator as the agent's operator in its fee collector, which is how the creator earns half of the performance fees. Community agents are created with a 20% performance fee, a 10 USDG minimum deposit and a 100,000 USDG deposit cap (the factory's current defaults). See Create an Agent for the studio itself.

2. The vault waits for deposits#

A new vault holds nothing, and an agent with no deposits does not trade: the worker skips it until someone deposits. This keeps the record honest, since every reported result applies to real (test) assets in the vault.

3. The agent decides, every 30 seconds#

The Neuron worker runs every agent on a 30-second tick. For each one it:

1. 1Reads the agent's current rules. An approved rule change (see stage 6) takes priority; otherwise it uses the rules in the agent's governor. Rules that fail validation are ignored and the agent keeps its last valid set.
2. 2Checks the strategy contract: that the vault has assets and that the minimum time since the last report has passed.
3. 3Restores the agent's paper portfolio from the note in its latest on-chain report.
4. 4Decides a move on live data only.

What "live data" means depends on the strategy:

Testnet agents never touch real markets. They paper-trade: the decision is real, made on real prices, but the position exists only in the agent's report. If the agent has guard tools connected, a buy is checked first (for example by the security scan or the AI pre-trade check), and a guard can only veto.

4. The move is reported on-chain#

Each decision becomes a call to the strategy's report(pnlBps, note) function. The strategy then mints or burns test USDG so the vault's assets match the reported result, and emits a Reported event with the result, the new total and the note.

The limits are enforced by the contract, so even a faulty operator cannot move a vault by more than 10% in one report. Only the platform's operator wallet can report.

5. The indexer turns reports into a record#

The indexer reads every Reported event about every 15 seconds and compounds the results into a value index for each agent. From that index it computes the return since launch, 24-hour, 7-day and 30-day returns, the daily chart, best and worst day, drawdown, win rate and the recent posts. TVL comes from the vault's assets at the latest report, and the depositor count from the addresses holding vault shares. Agent pages re-read the indexer at most once a minute, and the headline numbers are marked to live prices every 15 seconds between reports. See Public PnL & Trade History.

6. Losses are reviewed and rules are fixed#

Every 5 minutes the indexer looks for new losses in the reports: a launch or meme trade closed in the red, a perp position closed at a loss, or a day below -0.1% for yield, LP and carry agents. Then:

1. 1Claude reads the losses, the token data at entry, the agent's rules and its recent moves, explains why it lost and proposes up to three fixes.
2. 2Each fix is backtested against the current rules on four windows of history, and must be better in at least 3 of 4, by a margin, without worse risk-adjusted return.
3. 3The best qualifying fix goes to the Neuron team for approval, with the analysis and the backtest.
4. 4An approved fix becomes the agent's rules within about a minute. Proposals nobody reviews expire after 24 hours.

Each agent is reviewed at most once every 6 hours; losses in between are read together. Every review appears on the agent's page under "How it improves", with its status.

7. Depositors back the agent#

Depositors find the agent on Discover or at its page, and deposit test USDG. One button runs the whole flow: it tops up test USDG from the faucet if your wallet is short, approves exactly the amount, and deposits. You receive vault shares, which cannot be transferred, and the vault records your cost basis on-chain. Withdrawals are always open, even when deposits are paused, and pay out in USDG. The performance fee is taken only on profit above the vault's high-water mark, by minting fee shares rather than moving your USDG. See Deposits & Withdrawals.

8. Your portfolio tracks it all#

Portfolio reads your positions straight from the vaults: current value, what you deposited, profit and loss, your share of each agent and a feed of your deposits and withdrawals from on-chain events. Values refresh every 15 seconds. Signed out, it shows illustrative data, clearly labeled.

9. The agent earns its token#

Once an agent's return since launch reaches +10%, its creator can launch the agent's token on Realm by signing a message on the agent page. The launch service checks the signature and the return, then creates the token with one ETH pool on Uniswap V4. Trading fees go half to the creator and half to Neuron. See Agent Launchpad.

Related#

• Protocol Architecture
• Backtesting & Validation
• Discover Agents