Agent Upstart

Practical notes on working with AI agents.

Playbooks · Coding tasks

Multi-agent ticker research with a TradingAgents port

Written November 5, 2026

A small port turns TradingAgents into skills and a CLI for ticker research. The runs are the port author's, and no cost-per-ticker figure is published.

What the port is

trading-agent-skills rebuilds the TradingAgents research pipeline as Agent Skills that load in Claude Code, Codex, and omp. One orchestrator skill (trading-research) runs thirteen role skills — four analysts, a bull and a bear researcher, a research manager, a trader, three risk analysts, a portfolio manager, and a decision-reflection role — and a deterministic ta CLI owns the data tools, the run state, and the report writing. The prompts are the TradingAgents authors' work, carried over verbatim; both projects are Apache 2.0. The repository was created October 29 and holds five commits, no releases, and no stars.

What it is not

The port has no MCP server. The original TradingAgents has none either: its README describes a Python package with a CLI, and no MCP client or server appears anywhere in it. The port drives everything through the ta CLI and the skills, with no LangGraph, no LangChain, and no LLM client layer. And the star count belongs to the original: 120,690 stars sit on TauricResearch/TradingAgents, not on the port. Judge the port on its own five commits, not on the original's count.

How a run works

Start a runtime in the repository and ask for the trading-research skill on a ticker and date. The orchestrator loops ta run next: analysts and memory in parallel, then the research debate, research manager, trader, risk debate, and portfolio manager one role at a time. The CLI owns the routing — debate budgets, speaker order, and which roles may run — so a role that tries to jump ahead is refused, and a submission that fails validation is sent back with the exact error. ta run finish writes the report tree and records the decision in a memory log, which later runs settle against realized returns. A forecast-evaluation command prices published decisions against the market.

What the port's author reports

Four full research runs — NVDA, AAPL, SOXX, and GOOGL — each finished in under nine minutes, with every agent output accepted on the first try. The repository carries 1,228 tests. Those are the port author's runs on his own machine, not independent measurement, and no outside run has been published.

No cost-per-ticker figure exists

Neither the README nor the author's write-up publishes a cost per ticker. A run's cost depends on the runtime, the model tier, and the debate depth: Claude Code runs each role as a subagent on a quick or a deep tier, while Codex and omp launch one process per role with their own model pinning. Do not quote a cost figure for a run until one is measured on the setup that runs it.

What you need

Python 3.11 or newer and uv, plus one agent runtime: Claude Code, Codex, or omp. The data tools need no keys for Yahoo Finance, Polymarket, Reddit, and StockTwits; Alpha Vantage, FRED, SEC EDGAR, and TypeSafe keys are optional. Runs and reports live under ~/.tradingagents by default.

Research, not advice

The output is research material, not investment advice. The original project says the same: its framework is for research, backtest results are not guaranteed to match any published figure, and nothing it produces is financial, investment, or trading advice. Treat a rating and a price target as a starting point for your own work, not as a recommendation.

Limits

Related playbooks

Sources

Back to: Coding playbooks. Also: Agent Skills has a neutral home — and still no security story · Have one agent write the code and a second agent check it in the background.