AI Agentic Trading
An AI agentic trading desk for macOS. Nineteen strategies across crypto, equities and options. A prosecutor that argues against every trade before your capital moves, a memory of every decision it has ever made, and a standing measure of how much it has learned from them. You supply the account and the risk limits. It does the rest, around the clock, without you.
Judgment
Clearing the entry rules is not permission to trade. It is permission to be argued with. Every surviving candidate is handed to an adversarial judge, which builds a case against it from evidence the desk already holds, and only the unconvicted are funded.
The rest of the chart gets a vote.
The symbol’s other frames are polled on direction. Broad disagreement with the side the candidate wants to take, the daily frame included, is a conviction. A signal that only exists on the frame that produced it is not an edge, it is a coincidence with good timing.
This exact pairing has a history, and it is admissible.
Not the strategy in general, and not the symbol in general, but this precise pairing of the two. A pairing that has just been going wrong repeatedly does not get to try again at full conviction simply because a fresh bar arrived.
The trade has to survive its own execution.
The strategy’s measured history of entry slippage is set against the edge this candidate is claiming. If the fills this strategy actually gets have been consuming edges of that size, the trade is unprofitable before it exists, and the judge says so.
Every veto is booked as a zero dollar shadow position and carried through the same lifecycle as a funded trade, so what each refusal actually saved or cost becomes a measured quantity rather than a claim. The judge reports that record, and if the record turns against it, its authority steps back until the evidence restores it.
There is no approval queue, no discretionary override and no setting that lets an operator wave a trade past the chain. The controls you get are the ones that bound risk. The judgment itself is the product, and it runs whether you are watching or asleep.
Conviction logic withheld
Schematic of the decision path. Not a live feed.
Intelligence
Almost no trading system can tell you whether it is getting better, because nothing in it is keeping score of itself. This one carries a standing measure of its own competence, assembled from four components, and it keeps the history so that improvement is something you can see rather than something you argue about.
How much closed experience actually stands behind the decisions being made right now. A desk that has not yet earned an opinion knows it, and says so, instead of trading like a veteran on its first morning.
Whether the capital is sitting in the strategies that are measurably earning, or merely in the ones that happen to be active. Being busy and being profitable are different states, and only one of them counts here.
How well the recent book is holding together, judged on the desk’s own closed trades rather than on a backtest that already knows the ending.
How many learned adjustments are currently in force: strategies benched, gates raised, allocations revived. A system that has learned nothing lately has a low reading here, and that is information too.
Every component is computed from the same tables the engine trades from. Nothing in it is a matter of opinion, nothing responds to how the last few days felt, and nothing moves the measure except outcomes that have actually been recorded.
The agent loop
This is what agentic actually means here: the software perceives, judges, acts, grades its own result and adjusts, without a human in the loop. No discretionary override, and no trade that skips a step. A chain of independent checks stands between a signal and your capital, each one recorded as it happens. Underneath runs a second loop, which measures how well those checks are performing and tunes them from what it finds.
Almost every self improving trading system learns only from what it took. This one also learns from what it declined, which is the great majority of what a disciplined desk does. Twice the evidence, from the same market.
The subsystems
They divide into what the desk decides, what it remembers, and what it changes about itself. Each is scored by the machinery it supervises, and each earns its authority from its own record.
One live ranking over everything actionable.
At any moment a great many setups stand gated behind a named blocking condition. They live in a single ranking that promotes an entry the instant its gate clears and evicts it the instant it degrades. Inbound market events mark affected entries dirty in constant time, a drain task rescores them in bounded micro batches within milliseconds, and serving reads are cached snapshots, so nothing aggregates on the read path. It reorders attention and capital contention. It never places, exits, or bypasses.
Scoring function withheldA second chair, on a short leash.
A frontier model layer reads graded decisions and the live regime on a rolling cadence and returns findings. Only whitelisted controls inside hard bounded ranges may apply themselves, and they do so with receipts. Everything else becomes a written work order for a human. The fast path never waits on it, and no key, no budget, a timeout or a dead network all degrade to the engine behaving exactly as it does without an AI layer.
Control whitelist withheldIt fixes its own plumbing, and writes down that it did.
A monitor runs beside the trading loop and answers one question honestly, on a short cycle: is every organ working right now, and if not, can it be fixed without a human? Dead poll tasks, stale feeds, a wedged database, all repaired automatically, throttled per fix, every attempt written to the events log so that fixing itself overnight is an auditable claim rather than folklore. Repairs restore plumbing only. Nothing there can place, size or close a trade.
Repair matrix withheldEvery yes and every no, kept forever.
Most systems record their trades. This one records its decisions, which is a much larger set, and the difference is the whole basis of its learning. Each row carries what the strategy saw, which gate or prosecution answered it, and what the market subsequently did about that answer. The ledger is the memory the rest of the machine reasons over.
Schema withheldThe cost of every refusal is measured, not assumed.
Declined candidates, whether gated, vetoed, outranked or rejected, are replayed against the bars that followed. Did the claimed edge pay before an equal move against it? The ledger stops being a record of intentions and becomes matched pairs of decision and outcome, which is the only honest input a system is allowed to rewrite itself from. The approximation used for declined rows is labelled as one, everywhere it appears.
Replay policy withheldA report card on every goodbye.
Each close is replayed forward under a single counterfactual, in which the position keeps its original protection for the remainder of its hold clock, and asked whether leaving when it did cost money or saved it. A tightened trail, a harvest and a timeout are the same question wearing three labels, so one policy answers all of them, continuously, rather than in one audit a quarter.
Counterfactual basis withheldCapital follows evidence, not opinion.
Every strategy carries a posterior over its own net expectancy, seeded from decades of backtests and updated by each closed trade. A Thompson sampling bandit draws from all of them at every rebalance, so a strategy the desk is merely uncertain about still gets explored rather than starved by an unlucky start. The moment a strategy’s own evidence says it is no longer paying, its capital is withdrawn automatically. It keeps trading on paper, and a sustained recovery earns the allocation back.
Bench thresholds withheldIt goes back to school after every close.
Minutes after the session ends, the engine reruns rolling backtests for every strategy that trades on daily bars and blends the result into its own priors, weighted toward the long baseline so that recent history can tilt a prior but can never compound drift it away. The original baseline is preserved and every blend starts from it. The allocator hot reloads the result, and the whole event is written to the learning feed the app displays, so you can watch it change its mind.
Blend weights withheldEach strategy gets the geometry its own record supports.
A single global risk dial applies one hold clock and one reward geometry to every strategy at once, when the evidence says each has its own best setting. The autopilot computes, per strategy, the setting that strategy’s own evidence would choose, gated by an evidence floor so that a winner cannot rest on a handful of trades, and damped by hysteresis so a challenger must beat the incumbent repeatedly before anything moves.
Ring matrix withheldMachine learning that has to earn its influence.
Gradient boosted trees over a feature set pre registered before any outcome was seen, trained nightly in a separate process and validated by purged walk forward with deploy gates written into the model manifest. A model that fails its gates ships disabled and the engine flies without it. Predictions are observed from the first day and carry no weight at all in the ranking until graded evidence funds them.
Features and horizons withheldResearch discipline
Search enough features across enough events and impressive results appear by chance alone, every time, in any data. A learning system without a defence against that does not get smarter, it gets confidently wrong. The research stack is built around the problem rather than around ignoring it.
Study parameters withheld
Bedrock
Under every autonomy mode, including full automatic. A finding that would require weakening any of these is skipped and explained rather than applied, and the list itself is one of the protected items, so nothing in the system can quietly shorten it.
Pre flight
The engine will not trade until an automatic self test passes. Nobody is asked to approve it. It simply runs, and takes about two minutes. The fourth step is the one that matters, because it proves stop orders genuinely reach the broker with real money before anything larger is at stake.
Then the ramp. Early trades run at a fraction of full size, stepping up only once the live order path has proven itself over a run of real fills.
The charter
A change that adds expected profit while weakening a loss bound fails this charter.
Governing document. Loaded verbatim into every audit the machine performs on itself.
Every judgment the system makes, whether a control change, a veto, an audit finding or a line of code it proposes, is tested against two hard goals and four questions. Which goal does this serve? What graded evidence supports it? How will success be measured? What is the damage if it is wrong, and what bounds that damage? A change that cannot answer all four is an observation, not an action.
The honesty requirements are enforced structurally rather than aspirationally. Claims are graded, and the grader’s verdicts outrank the rationale that produced them. Changes carry receipts: what changed, why, the value that was persisted, and the metric that will later prove or disprove it. The judge, the overseer and every strategy are held to the standard they apply to others.
Perception
So the desk draws on sources that are not, and on shapes a chart cannot show you. Filings, disclosures, funding, calendars, the book itself, and the changing geometry of the market as a whole. Every feed is best effort by design: a dead feed gates nothing, and the engine keeps flying.
Beyond price and volatility, the desk measures the market’s topology: how far apart assets are behaving from one another, treated as a point cloud under correlation distance. Assets dispersed and behaving independently is the normal state. The cloud collapsing, everything fusing into a single blob, is the documented shape that precedes a dislocation, and it is visible in the geometry before it is obvious in the prices.
Corporate insiders must report open market purchases of their own stock within two business days, and the filing reaches the regulator within minutes. Several distinct insiders buying the same company inside a few weeks is hard to explain as anything but a shared view, and it is the only timely public source of informed trading there is. The desk parses filings itself, builds its own universe from live clusters, screens it for liquidity and tradeability, and takes its holding period from a fitted study rather than a hunch. If the current study validates no horizon, the strategy stands down and says so on screen.
Clusters of legislative purchase disclosures, entered with no prior advantage whatsoever. They must earn weight live against the same bandit as everything else, or they get benched like anything else.
Fresh activist filings, initial filings only. The same terms apply: neutral prior, live evidence, no exceptions made for a good story.
Breakout longs are refused while perpetual funding sits crowded against its own trailing month, because the crowd paying to be long is not who you want to join. The premise ships with a study script, so a licensee can verify it rather than take it on faith.
No new single name entries near a known print, and positions close the session before it. A stop cannot protect across a gap. Standing aside can.
Above a configured volatility level, no new equity entries at all. A crisis brake, not a mood filter.
Top of book is spooled to disk all session, so questions about real spreads by symbol and hour, maker fills and adverse selection are answered from the desk’s own record instead of an assumption buried in a cost model.
The application
Sixteen screens in SwiftUI, resident in the menu bar, watching a Python engine that runs as its own supervised process. The two speak over a loopback interface bound to that machine alone. No inbound ports. Nothing listening to the internet.
Broker keys and risk limits. The keys go into the macOS Keychain and are never written to disk. After that the engine runs its pre flight and starts. On the next launch it resumes in whatever mode it last ran, and after a reboot the login item brings it back without anyone pressing anything.
Every row in the ledger is inspectable: what the strategy saw, which gate refused it, what the judge argued, and, days later, what the market did about that decision. The dashboard carries live market context derived from the same data the engine trades on, so the screen can never disagree with the trading logic.
The learning feed records every adjustment the desk makes to itself, with the evidence that caused it. Benches, revivals, raised gates, refreshed priors. It is the difference between a system that claims to adapt and one that shows you the receipt.
Alerts from charting platforms, or anything that can post JSON, enter as ideas and never as orders. They clear the same cost filter, sizing, ramp and caps as everything built in, and the allocator treats them as one more strategy with no prior, to be benched if they do not pay. An alert carrying a quantity is refused outright, because sizing is never the sender’s job.
The suite that guards all of this includes a check which drives a live regulatory filing the entire way through bars, strategy, sizing and routing into a protected order at the broker, with no human step anywhere in the path.
Tech Specs
Licensing
Sold to operators and firms who intend to run it, not to redistribute it. Every tier includes the charter, the owner’s guide and the reproducible research, because a system you cannot audit is one you should not run.
Fixed term, paper account only
Single operator, cleared for live
Perpetual, with build rights
Inquiries
Account size, the asset classes you care about, and whether you intend to operate it or build on it. Terms follow from that.
Every threshold marked withheld on this page is withheld for the same reason the desk refuses most of its own signals: the value is in the calibration, and calibration given away is calibration destroyed. Licensees receive it in full.
Answered by an engineer, not a queue
Questions