An example desk: it runs itself, I still make every call.
Not a bot that trades. Five layers of machinery that watch the market and file reports, staffed by agents that read those reports, argue with each other, and hand me a ranked list. This is one trader's wiring — an example, not a template.
This is how one trader uses Anchoryx. It isn't the prescribed workflow, and none of it is required — the setup below reflects a discretionary options book on a multi-week horizon, and it took a lot of wrong turns to arrive at. A day-trading desk, a systematic one, or a book with a different risk tolerance would wire this differently and should. What may transfer is the shape: where the machinery stops, where judgment starts, and the rules that keep the two from blurring.
Signal rises, decisions descend
Five layers, sorted by one ruler: how much the system has already chewed the data before it reaches me. L1 is raw feed, L5 is judgment. The interesting line is between L4 and L5 — where the machinery stops and someone has to decide.
Raw external data with no opinion attached — regime state, sector flows, COT positioning, dealer gamma, institutional positioning.
Insight about one ticker or one position — price action, structures, spread analysis, the position modeler.
Aggregated across everything I hold — risk, concentration, hedge health, hedge simulation, reports.
Four agents that run on their own schedule and file findings: is the thesis still good, what regime are we in, is this setup heating up, what new ideas exist. Alongside them run deterministic checks — the risk register, tripwires, portfolio stress — that compute rather than reason.
Not another scheduled job. An agent session with standing orders: pull everything the layer below filed, resolve the contradictions, rank what matters, make the calls — then write the session down so tomorrow's pass remembers today's.
If any of this transfers, it's that line. The agents below it are witnesses — each testifies about its own lane and nothing else. The layer above is the judge.
That's why I stopped letting the macro agent recommend moves on individual positions. It declares posture only — risk budget, tilts, hedge stance — and the judge derives the per-name moves by crossing that posture against the book review and the risk numbers. A witness doing the judge's job is how I ended up with a desk that quietly contradicted itself.
The judgment loop
Four moves, every pass. The fourth is the one most setups skip.
Retrieve
Pull every agent's latest verdict at once — including which ones ran blind, and where two of them disagree.
Judge
Resolve the conflicts in context. Rank what needs attention. Make the call, name by name.
Save
Write the pass down — what it read, the posture it set, the ranked priorities, the actions recommended.
Recall
Open the next pass by reading the last few, so yesterday's unfinished business carries forward.
The judge never kicks off an agent run on its own — the agents fire on their own clock and it works with what's there. When I ask for a fresh run by hand, that's a different thing entirely, and it's fine.
Four agents, four questions
Each one answers exactly one question, fires on its own clock, and files to the same board. They never call each other — the record is the only place they meet, which is what makes them worth cross-examining.
Macro Sentinel
What regime are we in, and how should the book be leaning?
Portfolio-wide. It reads the macro tape, positioning, and the catalyst calendar, then writes and updates the scenarios everything else reasons against. It never calls a trim on an individual name — that is somebody else's lane.
- Reads
- Macro tape · positioning · catalysts
- Files
- Regime read · scenario updates
Thesis Sentinel
Is the thesis behind each position still good?
Book-wide. For every active anchor it goes back to primary data and re-derives the read — never reading what another agent concluded earlier — and grades whether the news accelerates, delays, or invalidates the reason I put the trade on.
- Reads
- Active anchors · news · price structure
- Files
- Per-anchor findings · invalidation flags
Setup Scout
Is this setup heating up yet?
Story-scoped. It watches the names I am tracking but have not entered, scores them against the scan profile the strategy specifies — momentum, breakout, pullback — and tells me when a watch item has actually become actionable. It does not rewrite the thesis.
- Reads
- Draft and pending stories · bars · flow
- Files
- Readiness scores · shortlists
Story Craft
What is worth looking at that I am not already in?
The idea generator. It scans the tradeable universe through a chosen lens and drafts candidate stories with a thesis, a structure, and an invalidation already attached — so a new idea arrives in the same shape as everything else on the desk.
- Reads
- Tradeable universe · research notes · themes
- Files
- Draft stories · candidate ideas
Running beside them are deterministic passes — the risk register, tripwire evaluation, portfolio stress, exposure reconciliation. They compute rather than reason, they can't talk themselves into a view, and when their arithmetic contradicts an agent's clean grade, the board flags the conflict and the arithmetic wins.
Three workspaces, three clock speeds
I run three, each with its own standing orders. Same tools underneath, completely different job — and, importantly, different permission to write.
Live
What needs attention right now, and what am I doing about it?
Pre-market, shortly after the open, and post-close it retrieves the overnight agent runs, resolves conflicts between them, ranks the book, and sets a posture — green, amber, or red — with a ranked priority list and a call on every position.
Then it saves the session. The next pass opens by reading the last one, so a note to watch a name into earnings on Monday is still live on Wednesday.
End of week
What did the week teach me, and what am I set up for on Monday?
Backward first: P&L and attribution across the book — which thesis paid, which one was right for the wrong reason, which stop was hugging structure instead of respecting it.
Then forward: research and catalyst staging. It reads the week that happened and the calendar ahead, and sets the frame every daily pass then operates inside.
Maintenance
Is the record still true?
The unglamorous one, and the reason the other two can be trusted. It works the data-concerns queue — bad marks, broken legs, mis-parsed structures — grooms the pre-trade brief pipeline, keeps theme membership honest, and retires scenarios and tripwires that have aged out.
Every read above is only as good as the book it reads. This one exists to make sure the book is not lying.
A theme's ticker set is the day-one risk cluster prior. Before there's enough realized correlation history to see a factor position built up over a few days, the risk engine leans on theme membership to know which of my names move together. So a misfiled position doesn't just look untidy — it quietly distorts the risk read. Which is why I have the organizer propose and never dispose.
The watch schedule
Scheduled runs are my alarm clock. They wake an agent so the work is waiting — and my day starts at judgment instead of at data collection.
| When | Watch | What it produces |
|---|---|---|
| Pre-open | Pre-market scan | Overnight moves against the book, gaps, what changed while I slept |
| Morning | Daily brief distilled | One regime read plus a capped set of per-name notes, deduped against the live feed |
| 08:30 ET | Newsletter reconcile | Outside ideas sorted against the book — risk flags first |
| 12:30 ET | Newsletter reconcile | Same pass on the midday letter |
| 15:30 ET | Pre-close risk sweep | Last look before the bell |
| After hours | Evening brief distilled | What changed, technical and flow reads, earnings hold-throughs |
| Friday | Book organizer | Theme membership swept, unfiled positions surfaced — proposed, never auto-applied |
External newsletters and market briefs arrive through their own bridges, get reconciled against what I hold, and land as research notes — never as instructions.
What actually makes it work
The architecture is replaceable. These constraints are the part I'd keep — every one of them exists because something went wrong without it.
Producers re-derive. Auditors grade.
When I ask for a book review, the agent goes back to primary data and derives the read live. I don't let it read what another agent concluded earlier. Otherwise I get an agent agreeing with an agent agreeing with an agent — a chain of confident restatements that never touched a price. Separate auditors are the only ones I let read a run's own output, and their job is to attack it.
Fact beats opinion.
The deterministic checks write to the same board as the reasoning agents. When a computed breach contradicts an agent's clean grade, the board flags the conflict and the arithmetic wins. An agent that has talked itself into "this looks fine" doesn't get to overrule a number.
Every agent stays in its lane.
Each one answers exactly one question and is structurally prevented from answering its neighbor's. The macro read doesn't call trims on individual names. The readiness check doesn't rewrite the thesis. Overlapping mandates gave me agents that hedged, deferred, and agreed with each other. Narrow ones give me something I can actually cross-examine.
Propose, don't dispose.
Anything that touches the structure of the book hands back a reviewable proposal and writes nothing until I approve it. Anything that feeds the risk engine gets a human in the loop by design, not by accident.
Outside opinion is context, never a directive.
Newsletters and market briefs come in through their own bridges, get reconciled against what I actually hold — names I'm in, names I'm tracking, names with no overlap — and land as research notes that color the next scan. A published buy call never becomes an action item on my desk.
Staleness is a real verdict.
Every run declares what it covered and how fresh its inputs were. A run that couldn't see part of the book is marked blind rather than quietly reporting on the part it could see. "I don't know" is a legitimate output; silent partial coverage is the failure mode that actually cost me money.
The desk remembers.
Every judgment pass is written down — what it read, what it concluded, what it recommended — and the next pass opens by reading the last few. Without this I didn't have a desk. I had a very articulate stranger who arrived every morning with no memory of the position I put on Tuesday.
One judge, and it is me.
The coordinator ranks, argues, resolves conflicts, and recommends. It arms nothing and enters nothing on its own. What the whole apparatus produces is a short, ranked, defensible list of things to think about — with the reasoning attached and the disagreements surfaced rather than smoothed over. The trade is still mine to put on.
If you're building your own
Copy the shape, not the schedule. Three of the mistakes I made are probably worth skipping.
I started with the clever agent. I'd start instead with the system of record — the thing that knows what you hold, what you planned, and why. Every agent above is a thin layer of reasoning over a database that takes itself seriously. The reasoning turned out to be the easy part.
I also drew my waterline far too late. Decide which layer is allowed to decide before you write a single prompt. Everything above that line needs memory and conflict resolution; everything below it needs to file its report and stop talking. Mine went wrong because the line wasn't drawn, so five components all thought they were in charge.
And build the auditor early. An agent grading its own homework passes every time — I didn't find out for a month.
Want the daily rhythm instead?
How the four cadences and the discipline loop play out trade by trade.
Or read how this desk is run as trading risk management software.
The machine watches. I decide.
That's one desk. Build the book your own agents reason against, and run it however it suits you.