Harold researches and originates. Mack executes, one dedicated agent per trade, inside the boundaries you set. Iris grades the whole book four times a day and corrects the drift. Every decision is logged with its reasoning, its evidence, and its order receipts.
Interface rendering. Illustrative data — not live positions.
Six layers. The last one feeds back into the second, which is what makes it a system rather than a pile of tools.
50+ sources across six timeframes. Price, options chains with greeks, 80+ indicators, alternative data, news, fundamentals. Automated gap detection and backfill, plus your own proprietary feeds.
Natural language in, structured analysis out. 45+ tools chained automatically, every number grounded against the tool output, every options contract validated against the live chain.
/hunt and the setups pipelines turn the universe into a ranked, priced slate. Promotion pushes a curated idea straight from chat into the execution platform as a live trade idea.
17 playbooks graded across 10 years and 4 market eras. The matrix picks the playbook and the risk tier; the grammar hands the executor its stop distance and time limit.
One agent thread per trade idea. Structured decisions, three-layer pre-trade validation, multi-broker routing, virtual stops where the venue has none, and full order receipts.
Four daily checkpoints grade the live book on adherence, expectancy and drift, then dispatch corrections — bounded by the same matrix that made the original call.
The strategist cannot execute. The executor cannot pick its own playbook. The reviewer cannot migrate a trade into a refuted strategy. These aren't prompt instructions — they're tool surfaces, and each agent is bound to its own.
Reads the question, chains the tools, reads the scorecard, and originates the trade idea with the playbook and risk tier the evidence supports.
One dedicated thread per trade idea, rebuilt fresh each cycle. Assesses the mission, decides, and submits — through exactly one order tool.
assess_mission is the single per-cycle readsubmit_decision is the only order or exit pathThe performance manager. Grades the fleet against the evidence, dispatches tuning, and escalates what it can't fix on its own.
Harold chains analytical tools on every query — deciding which to call, in what order, and how to synthesize the results. On Enterprise the output doesn't stop at a document: one conversation can produce a backtest, a validated multi-leg structure, and a trade idea promoted into the execution platform.
prepare builds the draft with no database write; promote validates and pushes the confirmed idea to your instance.Backtest: fold-mean R/DD +1.38, positive in 4/4 eras, 74% win rate, -6.4% worst-era drawdown. Tier solid, confidence confirmed for mega-cap daily.
Structure: sell 3× AAPL 245C monthly, buy 2× SPY 580P. Net credit $4,260, max risk $48,740 — inside the $50k bound.
Rendering of the research chat. Illustrative data.
Seventeen playbooks, each with a machine-readable grammar — entry expression, stop distance in ATR units, time-stop, trailing and scale-out behaviour, regime gate. Each one graded per asset segment, timeframe and security type.
The heavy statistics run offline in the research harness. What reaches the live path is a table: for this situation, these playbooks have proven edge, ranked, with a confidence label and a risk tier attached.
cost_dies and structurally excluded.| Segment | Tier | Fold-mean | Status |
|---|---|---|---|
| mid-cap | Solid | +1.42 | Locked |
| small-cap | Solid | +1.16 | Locked |
| mega-cap | Avoid | -0.08 | Gated off |
| crypto | Avoid | — | No chain edge |
Rendering of a style-matrix cell. Illustrative data.
Each active trade idea gets its own thread. The agent is rebuilt fresh each cycle so state can't quietly accumulate between decisions.
A single call returns the whole mission picture: live market state with the style's indicator pack, the regime stack across daily, 4-hour and 1-hour, open positions with P&L, working orders, remaining budget and buying-power coverage.
The agent evaluates that context against the resolved grammar — entry expression, stop distance, time-stop, exit rules — plus the thesis, any strategy intel Harold has distilled, and the setup-alignment cross-check that grounds its confidence.
A structured decision with reasoning and a confidence score. One tool submits orders and exits; everything else is read-only. Assist mode proposes and waits for approval, auto mode acts and then reports.
Approved orders pass three validation layers before reaching the broker. A post-decision reconciler then runs winner management and the deterministic options-exit grammar against what actually filled.
BUY_TO_CLOSE 3× AAPL 260320C245 @ $1.20 LMT · AlpacaRendering of an agent decision record. Illustrative data.
Autonomy without boundaries is just an unsupervised bot. Every dimension of what an agent may do is a configured limit, enforced below the model.
Most automated systems have no feedback loop — strategies drift, parameters go stale, and nobody notices until the drawdown. Iris closes that loop on a schedule.
Four times a day, in defined ET windows, Iris builds a scorecard for the live book: is each idea actually following its playbook, is expectancy tracking the evidence, are the monitor's flags being resolved. Then it dispatches corrections.
Rendering of the fleet review scorecard. Illustrative data.
Each trade idea is a self-contained execution environment with its own capital, broker connections, lifecycle rules and dedicated agent. Run dozens simultaneously, isolated from each other, on a per-user instance.
Every idea carries its own capital allocation, multi-leg positions across multiple symbols, per-leg broker routing, and lifecycle rules. Budget is tracked as allocated, used, reserved and available — and a batch pre-flight prices the whole basket before anything is submitted.
Route different legs of the same structure to different venues. Equities and options on one account, crypto spot on another, perpetuals on a third — each leg independently managed and reconciled.
Crypto and options venues often can't hold a native stop. The platform holds it instead — monitoring price on a tight loop and firing the exit itself when the level breaks.
Unwinding a covered position in the wrong order can leave you momentarily naked. The wave planner sequences it — coverage-consuming orders are staged behind the order that provides the coverage, and released only when it fills.
Wheels run themselves: cash-secured put, assignment, covered call, repeat — with a deterministic protective floor underneath the assigned-stock phase for when the thesis breaks.
The decision-time quote is stamped onto every order at submit, so fill quality is measured against what the agent actually saw — not against a price reconstructed after the fact.
Every order — whether a human placed it or an agent did — passes the same three independent checks. Options carry their own collateral math for cash-secured puts, covered calls and naked positions.
Ideas whose account lacks the buying power to run them are flagged capacity-blocked rather than left to fail at submission.
| Tier | Name | Max / order | Max daily loss |
|---|---|---|---|
| 1 | Capital Preservation | 10% | 1% |
| 2 | Conservative | 20% | 2% |
| 3 | Moderate | 33% | 3% |
| 4 | Aggressive | 50% | 5% |
| 5 | Degen | 100% | 15% |
Rendering of the risk dial. Values match the shipped tiers.
A permanent, per-user event journal. Agent decisions, order fills, stops and targets, assignments and exercises, pattern signals, outsized market moves, macro briefings and system alerts — typed, deduplicated and timestamped.
On top of it, a periodic analysis cycle: Harold reads the recent activity on a cadence that adapts to what you actually hold — every 15 minutes during regular hours or for round-the-clock crypto ideas, stretching to four hours overnight, and skipped entirely when nothing is live.
Rendering of the monitor feed. Illustrative data.
The market doesn't wait for your morning routine. Harold writes a briefing at every window that matters — from the Asia overnight session to the crypto evening wrap — and each one lands in your feed with the numbers already grounded.
Rendering of the briefing schedule and day-thesis pill.
Every tool returns structured data that can be chained, compared and built on across queries. Enterprise unlocks the execution and promotion categories on top.
Six timeframes, 80+ wide-format indicators, full options chains with greeks, volume-profile price structure, and symbol discovery across thousands of instruments. Proprietary feeds join the same interface.
Deterministic sims for equity grammar, modelled options premium and pairs, with realistic slippage. Parameter sweeps, regime analysis and portfolio risk diagnostics. Compute runs in an isolated sidecar, off the request path.
24 detectors with per-timeframe cooldowns, each signal carrying its backtested win rate and sample size. Detector performance itself is measurable, so a decaying signal can be identified rather than trusted.
NYSE trading calendar, earnings, options expiries and the macro schedule. Catalyst scans are batched into a single request so a wide hunt costs one call, not fifteen.
Financial statements, SEC filing search with full document retrieval, historical earnings, and on-chain crypto fundamentals.
Macro dashboards and FRED series, cross-source sentiment, live web search for breaking context, and prediction-market implied probabilities on political and economic events.
Number-level grounding on every response, contract validation on every options structure, research summaries and saved reports. Risk-profile awareness is enterprise-only.
Enterprise only. Read the book, promote a curated idea from chat, start and stop agents, approve or reject proposals, brief an agent with time-boxed intelligence, and raise flags from the autonomous cycle. Each surface binds a different subset — promotion never appears in an autonomous cycle.
The platform ships with market data out of the box. Enterprise clients connect proprietary feeds that become first-class citizens in every workflow — not a bolted-on side table.
Your data flows through the same pipeline as everything else: available in research, in backtesting, on charts, and inside the agent's per-cycle context. No separate system, no export and re-import.
Push data through authenticated REST endpoints or stream it over WebSocket for real-time ingestion.
Define schemas for proprietary signals, alternative datasets or internal research output.
Your feeds reach research, backtests, the toolset and the agents — exactly like built-in sources.
Read-only programmatic access to the whole data layer for your own models and systems.
Evidence-backed research workflows, walk-forward backtesting, and bounded autonomous execution for systematic and discretionary books alike.
Research infrastructure and order management that integrates with existing execution systems, with a complete typed audit trail for review.
Structured research, portfolio monitoring and compliance-ready analytical workflows, with per-team roles and separated permissions.
High-throughput data, real-time intelligence, and multi-strategy execution across equities, options, crypto and perps from one surface.
We'll walk you through research, origination, the evidence engine, agent management and live execution — on your symbols, with your scenarios.
Enterprise plans require verification of regulatory status or institutional affiliation.