The agent team and pipeline
Specialised Stratmill agents pass each strategy along a fixed sequence of stages with checks between them.
The agent team
The Research Lead writes hypotheses. The Strategy Developer codes them as NautilusTrader strategies. The QA Agent reviews the code for correctness. The Strategy Analyst reads backtest and optimisation results and decides what happens next. The Risk Officer audits a strategy's safety before promotion. The Portfolio Manager decides promotion and allocation across the whole book.
Supporting them, the backtest engine runs every backtest and optimisation with real fees, the data engineer keeps the market data catalogue complete, the Library Refiner proposes cited improvements from the Knowledge Library when a strategy is iterated, and the exporter ports strategies to other platforms.
Stages in order
A run moves through coding, verifying, backtesting, backtest review, optimising and analysing. From analysing it can iterate back to coding, go back to the Research Lead for a revised hypothesis, be abandoned, or move on to risk review, promotion and paper trading. Later stages are processed first so that verification and backtesting never starve behind a queue of new code.
Verification layers
Code is verified before it touches the full history. A resubmission identical to the previous iteration fails at once. A consistency check confirms the code trades the venue and instrument the hypothesis declares. Static analysis checks the structure and required methods and bans file, network and system access as well as ways of reading data from the future.
Synthetic tests then run the strategy on generated markets (uptrend, downtrend, flat, a volatility spike, zero volume and a price gap, plus missing supplementary data when the strategy uses it) in an isolated sandbox. A short sandbox backtest on recent real bars must trade at least once. Finally the QA Agent reviews the logic against the hypothesis: look-ahead, signal direction, sizing, venue rules and schedules. QA does not judge performance.
When verification fails
A failure returns the run to the developer with the finding. After 3 failed attempts the run is abandoned. A rule that never triggers, or a contract the developer cannot meet, goes back to the Research Lead for a revised hypothesis instead. Failures caused by the platform itself are retried and never counted against the strategy.
The two analyst decisions
After the first backtest the analyst decides optimise, iterate, revise the hypothesis or abandon, so hours of optimisation are only spent on ideas with a real signal. A viability guard checks the decision against the numbers: a strategy with enough trades and an average trade return above the venue's fee floor is not abandoned on a whim, and one below the fee floor is not sent to optimisation.
After optimisation the analyst decides promote, iterate, revise or abandon with the out-of-sample and holdout evidence in hand. Every decision is stored with its reasoning on the strategy page.
Fee floors per venue
The average return per trade must clear the cost of trading: 0.15% of notional on most perpetual futures, 0.10% on coin-margined and options-exchange perpetuals, 0.25% on most spot markets, 0.05% on US equities and spot gold, and 1% on prediction markets. A strategy below its floor is iterated or abandoned, because its edge would be paid away in fees.
Risk Officer review
Before promotion the Risk Officer audits the strategy's safety and must approve it explicitly. Code-enforced limits apply on top: leverage within your risk profile's cap, a defined risk control (a stop that risks a bounded share of equity per trade, or a signal exit with bounded position size), no critical issue and an acceptable risk score. The Risk Officer can also reject a gate waiver the analyst wrote. A rejection sends the run back to the developer.
Portfolio Manager and allocation
The Portfolio Manager stage re-checks every gate and risk limit in code, holds back strategies too correlated with what is already deployed, and decides whether to promote to paper trading. Capital is allocated by inverse volatility with a cap per strategy, new strategies start at equal weight, and incubating strategies get half weight. Moving from paper to live trading is never automatic.
Iteration limits and abandonment
A strategy is abandoned when its average trade return is below trading costs, when its out-of-sample performance fails, when it repeatedly fails verification, or when the analyst sees no path to viability. The reason is recorded and feeds back into the research agent so the same dead end is not repeated.
Platform runs and your runs
The platform runs its own research continuously; your workspace runs alongside it with reserved capacity and higher priority, so your idea is not queued behind platform work. You can pause, resume or stop your own runs at any time from the strategy page.