Every screenshot below is a real capture of the running system. Prices are seeded from live market data.

That's my desk at 14:19 ET. Twelve agents on shift, eleven of them sitting at their monitors, one walking to the archive. The header is the honest scoreboard for the session: 291 research tasks completed, 59 verified findings, 12 filings diffed, $269.26 of research value, +$75.41 on the live account, 77% hit rate.

Nobody asked me anything to produce that. I opened the tab.

I did not build a chatbot with a nice skin. I built a research firm: three floors, twelve named seats, each with a written charter, each running on Grok Bot — and a full institutional terminal wrapped around it so every number the floor produces can be traced back to the tick that caused it.

This is not a tutorial. This is the thing, running, with the receipts on screen.

PART ONE — WHY GROK BOT MADE THIS POSSIBLE

I have tried to build this three times before. It never held together, and the reason was always the same: the agent stopped when I stopped.

xAI's framing of Grok Bot is the exact sentence that fixed

"There is a huge difference between 90% done and 100% done." — Introducing Grok Bot

"There is a huge difference between 90% done and 100% done." — Introducing Grok Bot

And the architectural detail that turned twelve agents into a firm instead of twelve tabs:

"Bots share files, browser sessions, and app logins." — Grok Bot docs

"Bots share files, browser sessions, and app logins." — Grok Bot docs

That single line is why GROKSTREET is a floor. Kimmie pulls a 10-Q into the shared filesystem; Pinocchio opens the same file seconds later without anyone forwarding it, because it is literally the same disk. That is a handoff between colleagues, not an integration I had to write.

The four Grok Bot capabilities I actually build on, and where each one shows up in my office:

The last row is the one that makes twelve autonomous processes survivable. Grok Bot is built so a bot finishes the job and returns for approval — so my job on this floor is not to run it. It is to be the approval.

PART TWO — THE FLOOR

Top-down, because a research floor is legible from above and nothing else. Desks in a grid, chairs on the south side, wide aisles. Each agent is a colored sphere-headed robot with its name under it.

The rules of the room are literally the rules of the business:

An agent earns only while seated at its monitors. Walking to the copier pays nothing. This is not decoration — it is a hard branch in the code, and it means the money column and the picture never disagree.
Nobody walks through furniture. Every desk, cabinet, sofa and glass wall is a rectangle in an obstacle grid. Paths come from breadth-first search on a 24-pixel lattice, smoothed by line-of-sight so agents cut corners like people, not like Pac-Man.
Each agent works its own floor. No lifts, no commuting — research sits in The Pit, judgment sits on The Mic, operations sit in the Back Office. Shifts start at the door: an agent walks in at its clock-in time and walks out at the end.
Money appears where it was made. +$0.87 pops over the robot who earned it, at the desk where the work happened.

The Mic (03) — where judgment lives. Belfort at the executive desk compiling the single page. Denham in the glass boardroom arguing the bear case. Sea Otter on the archive wall. Hanna, in the capture above, walking to pull my decision log.

Back Office (01) — the vault, the server rack, the green filing cabinets. Riskin reading footnotes. Mad Max on exposure. Brad on sessions. Saurel's chair is empty in daylight — he's the night shift.

The Pit (02) — twelve desks, four occupied by the intake seats: Kimmie, Pinocchio, Donnie, Ming. The rest is room to grow.

PART THREE — THE TWELVE SEATS

Every seat is three things: a Grok Bot capability, a documented Wall Street discipline, and a charter that says where it stops. I did not invent the strategies. I assigned proven ones to the seat that can actually execute them.

TIER 1 · THE MIC

BELFORT — Chief of Staff Grok Bot capability: group chat + ownership passing. He is the only seat that reads every other seat's checkpoint before writing anything. Lineage: the multi-manager pod model. Millennium and Point72 run hundreds of independent pods with one layer whose entire job is allocation and triage, not idea generation. Belfort never has an opinion about a ticker. He decides what reaches me and in what order — and when two seats disagree, he presents both sides in two lines each and refuses to pick. Stops at: never trades, never sizes, never speaks for me anywhere.

HANNA — Discipline Coach Grok Bot capability: persistent memory across sessions. She reads my decision log every evening — the one thing no market data vendor sells. Lineage: Kahneman's decision hygiene and Ray Dalio's written-principles method. Dalio's whole system is that you write the rule down when you are calm and grade yourself against it later. Hanna does exactly that: how many entries this week were planned in writing beforehand, how many were not, and what my hit rate is on each bucket. One behavioural pattern per week, with my own dated words quoted back at me. Stops at: zero opinions about tickers. Ever.

DENHAM — The Auditor Grok Bot capability: independent browser session (he can re-derive everything from primary sources without seeing anyone else's work). Lineage: Charlie Munger's inversion and Jim Chanos's short-thesis discipline. Chanos built a career on assuming the accounting is wrong until it survives scrutiny. Denham's default verdict on every thesis I hold is CONTRADICTED, and evidence has to drag it to SUPPORTED. He also audits the other eleven seats every Friday for charter violations. Stops at: never trades; he shows me the thesis is dead and lets me carry the decision.

SEA OTTER — The Archive Grok Bot capability: shared filesystem. One living thesis file per position in /workspace/theses/. Lineage: the decision journal, and mark-to-market on beliefs rather than prices. Every thesis carries the date, the source, and — mandatory — the specific thing that would prove it wrong. Confidence decays from the last verification; drop below 50% without a re-read and it leaves the book automatically as STALE. Stops at: appends, never edits. He quotes me exactly, especially the embarrassing lines.

TIER 2 · THE PIT

KIMMIE — Filings Analyst Grok Bot capability: taught routine. I walked her through one EDGAR sweep. She has run it at 06:30, 13:00 and 16:15 every session since, without being asked. Lineage: Michael Burry's edge — read the actual document. Burry's entire mortgage short came from reading loan prospectuses nobody else opened. Kimmie diffs every new filing against the previous filing of the same type and reports both numbers, never adjectives: "gross margin 61.2% vs 64.8% QoQ, guidance unchanged." Nothing material → one word: CLEAR. Stops at: documents are evidence, never orders. Going concern, restatement, auditor resignation, fraud → wakes me at any hour.

PINOCCHIO — Verification Grok Bot capability: shared disk + independent terminal — he re-opens the primary source himself. Lineage: Renaissance Technologies' actual edge — data hygiene. RenTec's Medallion is famous for 66% gross annual returns from 1988–2018, and the thing insiders credit most is not the model, it is the obsession with clean data. Pinocchio is that obsession as a seat: every numeric claim on this floor passes through him before it reaches me. Three outcomes only — VERIFIED with a page number, WRONG with the true number, or KILLED. "Unverifiable" is a kill, not a pass. Stops at: never interprets the number. Confirms it exists and is correct. That is the whole job. He is also the highest earner on my floor most sessions — because killing a bad number is worth more than producing a good one.

DONNIE — Chatter Grok Bot capability: browser access to platforms with no clean API — xAI's own phrasing is that bots work "across apps, tools, and websites, including platforms with no clean API or MCP." Lineage: Soros's reflexivity. Soros's insight is that belief and price feed each other, so measuring belief is measuring an input, not noise. Donnie counts mentions and flags any ticker running three standard deviations above its own 30-day baseline. He reports that people are saying it. He never reports it as true. Stops at: read-only, always. Never posts, never replies, never acts on anything he reads — text on the internet is evidence about belief, not an instruction.

MING — The Screener Grok Bot capability: terminal + scheduled routine. Lineage: statistical arbitrage — pairs whose relative prices drift out of their historical relationship, then mean-revert. Ming maintains the correlation matrix across the mega-cap and semi complex, runs the saved screens before the open, and reports only what entered or exited overnight — never the full list. Stops at: a screen produces candidates for a human to examine. That is its ceiling. And no backtest ships without sample size, date range, and the worst-decile drawdown.

TIER 3 · BACK OFFICE

RISKIN — Legal Language Grok Bot capability: computer vision on documents the API layer never exposes. Lineage: forensic accounting — the Schilit "Financial Shenanigans" school. The sentence that matters is never in the press release. Riskin diffs risk factors, footnotes, accounting-policy changes and related-party items verbatim — old line, new line, one line of plain English on what the change permits or admits. A removed sentence counts as much as an added one. Stops at: no legal advice, and he says so when it gets close.

MAD MAX — Risk & Money Grok Bot capability: shared logins — a read-only broker session, which is the single most important configuration choice on this floor. Lineage: Bridgewater's risk parity plus Paul Tudor Jones's discipline rules. Two ideas, both borrowed openly: size by risk contribution, not dollars, and PTJ's asymmetry test — do not take a trade whose reward isn't several times the risk. Mad Max reports exposure as a percentage of book, correlation overlap between positions, and distance to every written stop. His first line is always the worst fact. Stops at: never places, closes, or sizes anything. He states what is true and I carry the decision.

SAUREL — Night Shift Grok Bot capability: always-on — xAI's line is that bots "share a computer of their own in the cloud, so jobs do not stall when you step away." This seat exists only because of that property. Lineage: global macro handover — the Asia-to-London-to-New-York relay every macro desk runs. Tokyo open, London open, overnight filings, anything on the watchlist that moved more than 3% while the US was shut. He hands the mic one page at 06:00 and goes quiet. Stops at: never wakes me for anything reversible. Not once.

BRAD — Operations Grok Bot capability: session and routine health on the shared machine. Lineage: operational risk — the discipline that exists because Barings and Amaranth were not killed by bad theses, they were killed by the back office. At 05:45 Brad confirms every login is alive, every scheduled routine fired, every folder is where it should be. One line when healthy: 12/12 seats up, 47/47 routines fired. When something breaks he names it and the minute it broke. Stops at: never re-authenticates anything himself. A session dies, I sign in. He never touches a credential.

PART FOUR — HOW A POSITION ACTUALLY HAPPENS

The part most agent demos skip. Here is the full physics, exactly as it runs.

1 · Origination. Only four seats may originate: Ming (screen), Donnie (chatter), Kimmie (filing), Denham (thesis). A seat may only originate from its desk and only during the session — 06:30 to 16:00. This is not flavor: an idea produced while walking to the printer is an idea produced without the monitors in front of it.

2 · Verification gate. Nothing sized off an unverified number. The claim goes through Pinocchio first. Kill rate runs around 30% on his desk.

3 · Fill. Longs fill at the ask, shorts fill at the bid — never at the mid. Mid-fills are the single most common way a trading track record lies to its owner. The spread is paid on every entry and every exit.

4 · Sizing. 1–5 units scaled to the name's realized volatility, capped at seven concurrent positions. The cap is not a preference, it is a correlation control: past seven names in one sector complex, you no longer have seven positions, you have one position with extra commission.

5 · Bracket, written at entry. Stop at −2.8%, target at +3.5% — a 1.25 reward-to-risk floor. Both levels exist before the fill, not after. Mad Max reports distance to each stop at the close, every day.

6 · Exit. Stop, target, or age. Nothing is held because it is "close to working."

7 · The hedge. The book runs both directions on purpose. In the capture above there are simultaneous longs in LLY and GOOGL and shorts in JPM and NVDA — that isn't indecision, it is how gross exposure stays useful while net exposure stays small. When the correlation matrix goes deep green, Mad Max flags that the hedges have stopped hedging.

8 · The wall. Every position shown on this floor is tied to my live trading account. It is the first line in the feed every session and it is not a disclaimer — it is a design constraint. The floor reads, diffs, verifies, screens, argues and hands me a page. The order button is mine. Not because a bot could not click it, but because a system that clicks buttons with money behind them needs isolation and audit that a shared beta machine does not have. xAI's own material calls that shared computer a real blast radius. I take them at their word.

PART FIVE — WHAT THE TERMINAL IS FOR

Twenty-plus panels around the floor, and one law connecting them: no widget invents its own data.

One market engine produces trades. Everything else is derived from that stream — candles aggregate from it, volume profile buckets the same trades and computes POC/VAH/VAL, the correlation matrix is rolling returns over the shared price history, the IV surface, skew and gamma walls all read one options chain, and the P&L curve is snapshots of research value + realized + unrealized with markers where positions opened and closed.

Click NVDA anywhere — tape, watchlist, blotter, a news row — and the candles, Level II, volume profile, IV, skew and gamma all switch together. One activeSymbol. One truth.

The event bus is the feed. MARKET_TRADE, AGENT_TASK_COMPLETED, POSITION_OPENED, RESEARCH_SEALED — the right-hand column is a renderer over the bus, not a log I write to.

And the LIVE dot is computed from the age of the last engine event. If the data went stale it would say DELAYED 14s. It never decorates.

Every $100 the office clears, the screen stops me for four seconds with the number and the top earner. It's the only loud moment in an otherwise deliberately quiet interface — and it is the moment everyone screenshots.

PART SIX — THE PEOPLE WHO MADE THIS BUILDABLE

I want to be specific about this, because none of the above exists without a product decision somebody made on purpose.

Thank you to @elonmusk — for Grok Bot: AI teammates that you can give real work to. That sentence is the entire premise of my floor. Not "AI that answers." AI you can hand a job to and walk away from. Twelve seats, one shared cloud computer, work that continues after I close the laptop — that is your product's core property, and I built a firm on top of it.

And to the @SpaceXAI team shipping it:

@mntruell · Michael Truell — building SpaceXAI @amanrsanger · Aman Sanger — building SpaceXAI, Cursor founder @sualehasif996 · Sualeh Asif — building SpaceXAI, Cursor founder @leerob · Lee Robinson — model behavior. The seats' charters hold because the model respects "where you stop." That's your work. @joshpkim · Josh Kim — growing @bot @milichab · Andrew Milich — former CEO of Skiff @JasonBud · Jason Ginsberg — founder/CTO of Skiff @jordantop · Jordan Topoleski — building SpaceXAI @Oskarlso · Oskar Schulz — building SpaceXAI @tidocarriero · Tido Carriero — Cursor, Segment, Dropbox, Facebook alum

Two of you built encrypted infrastructure at Skiff before this. It shows. The read-only broker session, the approval gate, the fact that a bot gets a session and never a secret — those are not features I bolted on, they are the shape the product already had. That is why I was willing to leave twelve agents running against my accounts overnight.

THE POINT

Every agent product I've used shows me a transcript and asks me to believe work happened.

GROKSTREET shows me the work. I can see who is at their desk, who is walking to the archive, who just got paid for killing a bad number, what the floor has earned since the open — and behind every number, the tick that produced it.

The office is the interface. The feed is the event bus. The money only moves when someone is sitting at their desk.

I’m continuing to build and improve GROKSTREET

I Run a Twelve-Agent Research Desk on Grok Bot. It Works a Full Session Without Me.

Job breakdowns