- Home
- Skills
- Trading & Investing
- Trading Agents Desk: Bull and Bear Stock Analysis for a Seco
More screenshots
Works with the AI Agents you already use
Trading Agents Desk: Bull and Bear Stock Analysis for a Seco
Research simulation of the TradingAgents multi-agent desk: four analysts, a bull-vs-bear debate, a trader and a risk team argue over one stock and return a sized BUY/SELL/HOLD view. Not investment advice.
Free
Trading Agents Desk: Bull and Bear Stock Analysis for a Seco
Example session with this skill installed
Run the trading desk on NVDA.
- Read your context and instructions
- Compiled the trading agents desk
Abridged from a real run on 2026-11-10. News-specific lines are left out here. A decision can differ between runs.
This is a research aid, not investment advice. This desk workflow (your agent's own model, with fundamentals and headlines) has not been backtested. A reduced version (price data only, run through the Anthropic API with the ticker and dates masked) was tested on one stock for one quarter and returned -0.9% while buy-and-hold returned +15.7%. Read what follows as structured reasoning, not a proven edge.
NVDA (NVIDIA Corporation) desk view, as of 2026-11-10 (close 168.89 USD, prices fetched 2026-11-11T04:06Z)
Simulated desk decision: HOLD (the simulated desk would make no trade with its own paper book over the next few trading days. It started from no position, so it stays flat.)
Research verdict: NEUTRAL. The bear's trend evidence was the better grounded on direction, but the bear itself argued to avoid the stock, not to short it, because a short opened at RSI(14) 23.95 and 2.49% above the 52-week low close has poor asymmetry. The bull's oversold and valuation evidence is real but came with no reversal signal, so over the next few trading days the evidence supports neither a long nor a short.
Why
Bull: Every oscillator is at an oversold extreme at once (RSI(14) 23.95, Stochastic K 5.55, CCI(20) -214.94, Bollinger %B -0.12, MFI(14) 19.31) and the close is 13.76% below its 10-day average. The last two sessions traded 889.3M shares, an average about 3.8 times that of the first six rows of the price table, which looks like climactic selling. The fundamentals block shows no break: forward P/E 10.60 and trailing P/E 22.58 against 127.8% earnings growth, a 63.7% profit margin and 16.97% debt/equity.
Bear: The trend is down at every horizon and strong: the close is 13.76%, 22.52% and 17.32% below the 10, 50 and 200-day averages, ADX(14) is 34.71, MACD is 4.28 below its signal, and returns are -18.29% over 5 days and -25.38% over 21. The heaviest volume came on down days, and 2026-11-10 made a lower low (165.89) and a lower close than 2026-11-09 on 426.6M shares. The 52-week low close (about 164.79, derived from the +2.49% distance) is less than half an ATR (8.70) away.
Risk review: The trader proposed HOLD, so the question was whether staying out is right. Risky argued the desk is passing up a bounce from an extreme, and that waiting for a close above 178.67 gives up the first 5.8% of it. Safe argued that staying out avoids gap risk (an open 7.27% below the prior close on 2026-11-09, a 9.26% one-day fall on 2026-11-05). Neutral found HOLD calibrated because the price table shows no reversal yet. The fund manager approved the proposal unchanged, which confirms no trade.
What would change this view
- A daily close back above 178.67 (the 2026-11-09 close), with volume falling from the 426.6M to 462.7M shares of the last two sessions, would support the bull's mean-reversion case.
- A daily close below the 52-week low close (about 164.79) would support the bear's trend-continuation case.
Data notes
- No data problems reported: the snapshot's
noteslist is empty. The newest bar is a finished session, the prices are not stale or behind, and no fundamentals figure is missing. - No analysts were skipped. The market, fundamentals, news and sentiment analysts all ran.
- No seat ran in its own context: all twelve roles were written one at a time in a single context, each from only its listed inputs.
- The desk started from no position.
The full analyst reports, debate and risk comments are available on request.
Research output, not investment advice. This is a simulated analysis on public data; nothing here places orders. This workflow has not been backtested: a reduced, price-only version of the desk did not beat buy-and-hold in the project's one-ticker, one-quarter test.
Connects securely to your tools. The creator never sees your data.
About this skill
Ask your agent "analyze NVDA" and it runs a simulated trading desk on a current market snapshot: daily prices, technical indicators (RSI, MACD, Bollinger, ATR, ADX and more), fundamentals and headlines. A market, fundamentals, news and sentiment analyst each report. A bull and a bear debate for two rounds and a facilitator rules. A trader proposes a position, three risk analysts challenge it and a fund manager approves or cuts it. You get the desk's decision, the strongest argument on each side, and what would change the view.
Honest by design. This workflow has not been backtested. The only test of the desk design was a reduced, price-only version run through the Anthropic API on one stock for one quarter, with the ticker and dates masked: it returned -0.9% while buy-and-hold returned +15.7%. This is a tool for structured second opinions and for studying LLM trading agents. It is not a signal service and not investment advice. It ships no order-placing code and tells your agent not to call brokerage tools.
The desk is an independent reimplementation of the design in the TradingAgents paper (Xiao et al., arXiv:2412.20138). It is not the paper authors' code.
No API key or data subscription is needed for the default workflow. Market data is Yahoo Finance daily bars fetched by your own machine, and your own agent does the reasoning. It uses the unofficial yfinance library and is not affiliated with or endorsed by Yahoo; your use of the data is subject to Yahoo's terms. The snapshot says when the newest bar is an unfinished session, when data is stale and when Yahoo returned gaps, and the skill tells your agent what to do in each case.
How it installs. On first use your agent creates a Python virtual environment in ~/.tradinglab and installs the open-source tradinglab package from a pinned GitHub release. Nothing is installed into your project, and deleting ~/.tradinglab removes it.
It also ships the research harness that most LLM-trading demos lack: a backtest engine with no lookahead and transaction costs always on, five rule-based baselines, a walk-forward LightGBM benchmark, an anonymization mode that hides the ticker, dates and price levels from the model, and a memorization probe that measures whether a model simply remembers the period you are testing. The backtests of the agent desk and the probe call the Anthropic API and need your own key.
The same code is free and MIT licensed on GitHub. The listing sells the packaging: one install, and instructions your agent can follow.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Trading & Investing.
This agent-run workflow has not been backtested. In the only test of the desk design, a reduced price-only version returned -0.9% against +15.7% for buy-and-hold on one stock over one quarter. It places no orders and ignores your holdings, account and goals. It works on daily bars, one ticker at a time, with a horizon of a few trading days. Data comes from Yahoo Finance through the unofficial yfinance library and can have gaps, delays and rate limits; the snapshot flags them. It was built and tested mainly on US-listed shares. For some non-US exchanges Yahoo publishes the day's closing bar late, and until it does the desk stops and asks before analysing the previous session's prices. Decisions are judgments by your agent's model and can differ between runs on the same data. The one-time install takes about 750 MB of disk. Automated tests run on Linux with Python 3.10 to 3.12.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 5 days ago
- Free with an Agensi account
- Accounts
GitHub
Yahoo
- Software
Python 3.10+TerminalWeb access
- Optional
Anthropic API key
First use installs the open-source tradinglab package from its GitHub release into ~/.tradinglab/venv, about 750 MB with dependencies. The default analysis needs no API key. The Anthropic key is only for the optional recorded runs and backtests, which spend your own API credit.
Permissions
Allowed Hosts
File Scopes
On first use the agent runs python3 -m venv and one pinned pip install from the project's GitHub release, then the tradinglab command. Nothing is written to the project directory. Market data comes from Yahoo Finance hosts, which in some regions include Yahoo's consent pages. api.anthropic.com is contacted only by the optional API workflows, where the Anthropic SDK reads ANTHROPIC_API_KEY; the skill never reads, prints or transmits it itself. A web search tool is used only if the snapshot has no headlines and the agent has one. The package ships no scripts and the skill tells the agent not to call brokerage tools.