<- all tokdocs
2026.08.31@ray_fu

The Repo Behind This Clip Was Published 20 Months Earlier, Has 109,820 Stars, and Never Places a Trade

Finance
Watch on TikTok

View on TikTok ->

This is a 40-second vertical clip at 1080x1920, posted 2026-08-31 by @ray_fu (channel nickname "Ray Fu"), carrying 144,800 views, 10,600 likes, 2,160 comments, 920 reposts and 8,305 saves at the time I captured it on 2026-10-05, scored to "Every Living Breathing Moment" by Clavier. I read 12 of the 20 extracted frames and the full 192-word transcript. The clip opens on a title card reading "How to build" / "a trading firm" / "with AI Agents" with "trading firm" and "AI Agents" set in yellow, over a caption that reads "that someone just built-" / "an entire" and then "Wall Street trading firm" / "with AI agents". Behind the title sits a document screenshot headed "Analyst Team" whose bullets read "Fundamentals Analyst: Evaluates company financials and performance metrics, identifying intrinsic values and potential red flags.", "Sentiment Analyst: Aggregates news headlines, StockTwits, and Reddit chatter into a single sentiment read to gauge short-term market mood.", "News Analyst: Monitors global news and macroeconomic indicators, interpreting the impact of events on market conditions." and "Technical Analyst: Utilizes technical indicators (like MACD and RSI) to detect trading patterns and forecast price movements.", with a "Researcher Team" heading cut off below. A second overlay shows a chat app with a chart labelled "AAPL" / "$238.42" / "+4.23 (+1.34%)" and cards for "Macro Strategist", "Risk Manager" and "Options Flow" above a "Message Claude..." box. At 0:04 the app fills the frame with a ticker tape reading "AAPL ▲ 1.34% MSFT ▼ 0.18% NVDA ▲ 3.10% TSLA ▲ 3.39% AMZN ▲ 0.87% GOOGL ▲ 0.25% META ▼ 0.52% • LIVE", a header "6 desks analyzing AAPL" with a counter "0/6 · 0.8s", and desk cards for "Technicals", "Sentiment Desk", "Risk Manager" and "Options Flow". At 0:08 a user bubble asks "Should I go long on Orbit Dynamics (ORBT) before earnings?" and the assistant replies "On it — pulling in four specialist agents to look at this from every angle before we talk trade." under a divider reading "4 AGENTS WORKING IN PARALLEL", with cards for "Fundamentals Analyst / Reading 10-Q filings...", "News Analyst / Scanning breaking headlines...", "Social Sentiment Analyst / Tracking chatter across social..." and "Price Action Analyst / Mapping price & volume...". At 0:12 the Social Sentiment card returns "Done" with "Bullish mentions up 310% in 48 hours. Sentiment score: 78/100 — though the spike is concentrated in retail-heavy channels." At 0:16 a debate pane shows "Bear agent: A CFO selling shares right before earnings is exactly the detail retail sentiment always misses.", "Bull agent: One scheduled sale under a 10b5-1 plan is routine — it's not a red flag on its own." and "Bear agent: Routine or not, it's landing next to compressing margins and an RSI of 71. That's how a good story turns into a crowded trade.", closing on a banner reading "⚠ HIDDEN RISK FLAGGED / Insider selling, margin compression, and overbought price action are converging right befor". The composer below reads "Reply to Claude..." and the footer reads "Claude can make mistakes. Please double-check responses." At 0:18 a left sidebar labelled "Claude" shows "New chat" and a "RECENTS" list reading "ORBT execution & sign-off", "Orbit Dynamics (ORBT)", "Q3 earnings watchlist", "Portfolio rebalance ideas" and "Macro outlook notes", with a second ticker tape reading "VTRA $76.01 ▲ 0.94% LUMN $15.88 ▼ 1.95% ARCO $301.50 ▲ 1.60% PXLR $54.57 ▼ 1.50% ORBT $89.42 ▲ 1.85% NOVA $143.12 ▼". The trade ticket beside it reads "Side BUY", "Ticker ORBT", "Qty 640 sh", "Order type Market", "Entry $89.50", "Stop $85.80", "Target $101.00", "Size 3.4% of portfolio", "Est. risk 1.2% of AUM". At 0:22 a red block reads "Risk Management Agent / Vetoed — a 3.4% position exceeds the 2.5% single-name limit, and 0.82 correlation with the existing NOVA position pushes sector risk over budget." followed by "Trader Agent — Plan v2 (revised) / Revising plan (v2)...". At 0:24 the revised ticket reads "Order type Limit @ $89.50", "Entry $89.50", "Stop $86.90", "Target $99.50", "Size 2.1% of portfolio", "Est. risk 0.7% of AUM", approved with "Risk Management Agent / Approved — the revised 2.1% sizing fits inside both the single-name and sector correlation limits, using 0.7% of the risk budget.", then "Portfolio Manager Agent / Reviewing final ticket...". At 0:26 a ledger lists "#41 HELX LONG Feb 18 +4.0% +5.1% A", "#47 NOVA SHORT Mar 12 -6.0% +9.2% F", "#55 QNTM LONG Apr 30 +3.0% -1.4% D", "#63 ORBT LONG Jun 21 +8.0% +11.3% A" and "#71 VTRA SHORT Aug 09 -3.0% -4.2% A", under a header "MEMORY AGENT REVIEWING THE MISS" and a card reading "Memory Agent — Trade #47 (NOVA, short)" with "PREDICTED MOVE -6.0%", "ACTUAL MOVE +9.2%", "TRADE P&L -9.2%" and "Lesson: the short ignored strong sector-wide momentum going into earnings. Added a standing rule so future shorts check sector beta first.", closing on a cut-off "PLAYBOOK UPDATE" header. The last 8 seconds are a talking head with captions "i put together a complete setup" and "just like the video and comment".

The system is TradingAgents, and it was published 20 months before this video, not "just built"

The clip's first line is "Most people don't realize that someone just built an entire Wall Street trading firm with AI agents." The word "just" is wrong by roughly 20 months. The Analyst Team bullets pasted on screen at 0:00 are copied nearly verbatim from the README of TauricResearch/TradingAgents, which describes the same four roles in the same order: Fundamentals Analyst, Sentiment Analyst, News Analyst, Technical Analyst. The underlying paper, TradingAgents: Multi-Agents LLM Financial Trading Framework by Yijia Xiao, Edward Sun, Di Luo and Wei Wang, was submitted to arXiv as v1 on 28 December 2024 and has been revised through v7 on 3 June 2025. The GitHub repository was created on 2024-12-28T03:31:08Z, the same day. This video was posted on 2026-08-31.

The repo is not obscure. Reading the GitHub REST API directly on 2026-10-05, TauricResearch/TradingAgents reports 109,820 stars, 21,114 forks, 822 watchers and 85 open issues, licensed Apache-2.0, last pushed 2026-10-03T21:30:10Z. The clip's framing that "most people don't realize" this exists describes one of the most-starred repositories on GitHub.

The agent roster the narrator describes is accurate to the source. The README and the paper both specify four analysts working in parallel, a researcher team of bullish and bearish agents running a structured debate, a trader agent, a risk management team, and a portfolio manager who approves or rejects the proposal. The clip's sequence matches the paper's five-stage pipeline. The persistent-memory claim is also real: the repo ships a memory module, and the paper describes agents reflecting on past decisions against realised market outcomes.

The on-screen app is a mockup, and the ticker it analyses belongs to a different company

Nothing on screen is a shipped product. The interface carries Claude's chrome, including the "Reply to Claude..." composer and the footer "Claude can make mistakes. Please double-check responses.", but Claude has no live market-data ticker tape, no "6 desks analyzing AAPL" panel, and no trade-ticket UI. Anthropic's own product page for Claude for Financial Services lists market-data connectors from LSEG, FactSet, S&P Global, Morningstar and PitchBook, and frames the product around client onboarding, meeting prep, portfolio reviews and compliance reviews with "review steps and an audit trail built in". It does not place orders. The animated desk cards, the veto banner and the order ticket in this clip are a designed demo, not a screen recording of software a viewer can open.

The ticker symbols give it away. The clip's central position is "Orbit Dynamics (ORBT)", quoted at $89.42 on the tape and bought at $89.50. ORBT is a real symbol, but it belongs to Orbit International Corp., a Hauppauge, New York electronics and power-unit manufacturer founded in 1957 that trades over the counter. There is no US-listed issuer called Orbit Dynamics. HELX, QNTM, VTRA and PXLR likewise do not match the companies implied by the ledger. The AAPL quote of "$238.42 +4.23 (+1.34%)" and the second tape's "NOVA $143.12" are set dressing.

One smaller on-screen error: the burned-in caption at 0:04 reads "you give the system a single stock ticket". The word is ticker.

The open-source framework returns a recommendation; it does not execute an order

This is the clip's largest substantive gap. The video shows a filled order ticket with "Side BUY", "Qty 640 sh", "Order type Market" and then "Limit @ $89.50", a stop, a target and a position size expressed as a percentage of AUM, and the narration says "a trader agent plans the execution". A viewer reasonably infers that cloning the repo gets them a system that trades.

It does not. The TradingAgents README describes a run as producing a buy, sell or hold recommendation with a confidence rating, evaluated against a simulated exchange. The package tree at tradingagents/ contains agents, dataflows, graph, llm_clients, memory, backtest.py, portfolio.py, reporting.py and report_html.py. There is no broker adapter, no order-routing module and no credentialed execution path. Connecting it to a live account at Alpaca, Interactive Brokers or anywhere else is work the viewer would have to do themselves, and it is the part where every real consequence lives. Third-party forks such as AlpacaTradingAgent exist precisely because the upstream project does not do this.

The data side is also narrower than "the entire AI trading floor goes to work" suggests. The README names Yahoo Finance and Alpha Vantage for prices, SEC EDGAR for US fundamentals, FRED for macro series, and StockTwits, Reddit and news aggregation for sentiment. Those are the same free and freemium feeds any retail screener uses, not institutional order-flow or options-flow data. The clip's "Options Flow" desk card has no counterpart in the open-source project.

The only published returns come from a 3-month backtest the authors themselves qualified

The video quotes no performance numbers, which is to its credit. Viewers who go looking will find them, so they are worth stating with their caveats attached. The arXiv HTML version of the paper reports a backtest window of 1 January to 29 March 2024 against five baselines (Buy and Hold, MACD, KDJ+RSI, ZMR, SMA), using gpt-4o-mini and gpt-4o for quick-thinking roles and o1-preview for reasoning-heavy roles. On AAPL it reports 26.62% cumulative return, 30.5% annualised, a Sharpe ratio of 8.21 and 0.91% maximum drawdown. On GOOGL, 24.36% / 27.58% / 6.39 / 1.69%. On AMZN, 23.21% / 24.90% / 5.60 / 2.11%.

Those Sharpe ratios are not plausible as a durable expectation, and the authors say so. The paper states the benchmark ran only three months "due to intensive LLM and tool use" and notes that there were few pullbacks during that period. A Sharpe above 6 measured over roughly 60 trading days on three mega-cap technology stocks during a strong rally is a demonstration of a pipeline, not evidence of an edge.

There is no audited track record for this system. No GIPS-compliant composite, no third-party performance verification, no filed Form ADV Part 2A with performance disclosures. Any return figure anyone attaches to this framework, including the paper's own, is unverifiable in the sense that matters: nobody independent has watched it trade real money through a full cycle. I also found a replication repository, M0-AR/trading-agents-falsification-2026, claiming the AAPL result does not reproduce and that returns go negative at 50bps of transaction cost. I checked it against the GitHub API on 2026-10-05 and it reports 0 stars, created 2026-10-05T10:30:21Z, pushed four minutes later. It is a single unreviewed repository published the same day I read it, so I am citing it as existing, not as established. The paper's own limitation paragraph is the stronger caveat and it does not depend on anyone else's work.

Running an actual trading firm triggers registration the clip never mentions

The video's title is "How to build a trading firm with AI Agents". The word "firm" is doing a lot of unexamined work, and the clip's call to action is "comment firm and I'll send it to you". Nothing in the 40 seconds mentions a single licensing requirement. Here is what the regulators' own pages say.

If you only trade your own money, you are largely outside this. The SEC's Guide to Broker-Dealer Registration states that "the definition of 'dealer' does not include a 'trader,' that is, a person who buys and sells securities for his or her own account, either individually or in a fiduciary capacity, but not as part of a regular business", and that "individuals who buy and sell securities for themselves generally are considered traders and not dealers". Running this framework against your own brokerage account is a personal trading decision, not a regulated business.

The moment anyone else's money or advice is involved, that changes. The SEC's small-business page on Investment Advisers defines an investment adviser as "a natural person or company that, for compensation, is engaged in the business of providing advice to others or issuing reports or analyses about securities", and says such a person "is required to register with the SEC or applicable state securities regulators as a registered investment adviser (RIA) unless they are exempt from applicable registration requirements". The split is by regulatory assets under management: under $25 million is a small adviser registering with states, $25 million to $100 million is mid-sized, and SEC registration is permitted at $100 million and required at $110 million, with de-registration below $90 million. Selling signals from an agent swarm for a fee is squarely inside "issuing reports or analyses about securities" for compensation.

On the broker-dealer side, the same SEC guide quotes Section 15(a)(1), which "generally makes it unlawful for any broker or dealer" to use interstate commerce to "effect any transactions in, or to induce or attempt to induce the purchase or sale of, any security" unless registered under Section 15(b). FINRA Rule 1210 then requires that "each person engaged in the investment banking or securities business of a member shall be registered with FINRA as a representative or principal in each category of registration appropriate to his or her functions and responsibilities". The proprietary-trading carve-out has narrowed, too: the SEC's 2023 amendments to Rule 15b9-1 removed the exemption that let registered dealers do unlimited proprietary trading without joining a national securities association.

If the strategy touches futures, options on futures, retail forex or swaps rather than cash equities, a different regulator applies. The NFA's Who Has to Register page defines a commodity trading advisor as a person or organisation that, "for compensation or profit, advises others, directly or indirectly, as to the value of or the advisability of buying or selling futures contracts, commodity options, retail off-exchange forex contracts or swaps", and a commodity pool operator as one who "operates a commodity pool and solicits funds for that commodity pool".

Finally, the model vendor imposes its own condition. Anthropic's Usage Policy classifies "use cases related to financial decisions, including investment advice" as high-risk, and requires that "a qualified professional in that field must review the content or decision prior to dissemination or finalization", plus disclosure to end users that AI helped produce the recommendation. A fully autonomous agent chain that signs off on its own tickets and ships them to someone else is not a configuration the policy contemplates. The project itself agrees: the README states that "TradingAgents framework is designed for research purposes" and "is not intended as financial, investment, or trading advice".

Key Takeaways

  • Correction: "Someone just built" is off by about 20 months. The arXiv paper went up 28 December 2024 and the GitHub repo was created the same day; this video posted 2026-08-31.
  • Correction: The clip implies a system that executes trades. The upstream TradingAgents package has no broker adapter and no order-routing module. It returns a buy, sell or hold recommendation evaluated against a simulated exchange.
  • Correction: "Orbit Dynamics (ORBT)" does not exist. ORBT is the real OTC symbol for Orbit International Corp., a 1957-founded electronics manufacturer in Hauppauge, New York. The demo attached an invented company name to a live ticker.
  • Partial correction: The interface is a mockup wearing Claude's chrome. Claude has no live ticker tape, no desk panel and no order ticket. Claude for Financial Services connects LSEG, FactSet, S&P Global, Morningstar and PitchBook data for analysis and review workflows, and does not place orders.
  • Verified: The agent architecture the narrator describes is accurate to the source. Four parallel analysts, a bull-versus-bear researcher debate, a trader, a risk management team with veto power, a portfolio manager sign-off, and a reflective memory module are all in the paper and the repo.
  • Verified: TauricResearch/TradingAgents reports 109,820 stars, 21,114 forks and 822 watchers, Apache-2.0 licensed, read from the GitHub REST API on 2026-10-05.
  • Unverifiable: No audited track record exists for this framework. The published numbers (AAPL 26.62% cumulative, Sharpe 8.21) come from a single 3-month backtest, 1 January to 29 March 2024, on three mega-cap stocks, which the authors explicitly qualified by noting the short window and the absence of pullbacks.
  • Unstated cost: The clip never mentions LLM spend. The paper's own configuration used o1-preview for reasoning-heavy roles, and the authors cite "intensive LLM and tool use" as the reason they could only afford three months of backtesting. Per-ticker, per-day cost is a real operating expense that the 40 seconds skips entirely.
  • Unstated requirement: Not one licensing obligation appears in the video. Trading your own account sits inside the SEC's trader exception, but advising others for compensation pulls in Investment Advisers Act registration at the state or SEC level, effecting transactions pulls in Section 15(a)(1) and FINRA Rule 1210, futures and forex pull in CFTC/NFA registration as a CTA or CPO, and Anthropic's own policy requires a qualified professional to review financial recommendations before they reach anyone.
  • Note: The call to action is lead capture. "Comment firm and I'll send it to you" routes viewers to DMs for something that is a public, Apache-2.0 licensed repository anyone can clone without commenting.

Resources

Published August 31, 2026. Writeup generated from a favorited TikTok.