The rise of multi-agent investing systems

In December 2024 a small group called Tauric Research published a paper and a repository with the same name: TradingAgents. The idea was simple to explain and, it turned out, very easy to like. Instead of asking one language model "should I buy this stock?", you model a trading firm. Analysts gather evidence. Researchers argue the bull and the bear case. A trader proposes a position. A risk team pushes back. A manager signs off.

Less than two years later the repository has 110,565 stars and is still gaining about +889 a week. A Chinese-language fork, TradingAgents-CN, has 32,243 of its own. The pattern has been copied, extended and, in a few cases, quietly replaced by something more conservative. This article is about that pattern: where it came from, how the main projects implement it, and what the evidence for it actually looks like.

TradingAgents

Stars

110,565

Stars gained · 7d

+889

Stars gained · 30d

+6,452

Last push

8 days ago

License

Apache-2.0

AI Hedge Fund

Stars

63,932

Stars gained · 7d

+115

Stars gained · 30d

+780

Last push

9 days ago

License

MIT

FinRobot

Stars

8,202

Stars gained · 7d

+68

Stars gained · 30d

+256

Last push

13 days ago

License

Apache-2.0

RD-Agent

Stars

14,893

Stars gained · 7d

+57

Stars gained · 30d

+334

Last push

5 days ago

License

MIT

TradingAgents-CN

Stars

32,243

Stars gained · 7d

+119

Stars gained · 30d

+620

Last push

19 days ago

License

Other

Project	Stars	7d	30d	License	Last push
TradingAgents	110,565	+889	+6,452	Apache-2.0	8 days ago
AI Hedge Fund	63,932	+115	+780	MIT	9 days ago
FinRobot	8,202	+68	+256	Apache-2.0	13 days ago
RD-Agent	14,893	+57	+334	MIT	5 days ago
TradingAgents-CN	32,243	+119	+620	Other	19 days ago

The org chart as an architecture

TradingAgents' pipeline, in the project's own description, runs as a LangGraph graph: a parallel analyst team (fundamentals, sentiment, news, technical) feeds a bull and a bear researcher who debate; a trader turns the debate into a proposal; a risk-management team reviews it; a portfolio manager makes the final call. The framework lets "quick-think" and "deep-think" steps run on different models, so a cheap model can summarise news while an expensive one argues the thesis.

AI Hedge Fund takes the same idea in a more playful direction. Its agents are investor personas assembled into "funds" with their own strategies, capital and cadence, stepped through completed market sessions with an approval step and a hash-chained ledger. The project is explicit that it is educational and never places real trades, which has not stopped it collecting 63,932 stars.

FinRobot, from the AI4Finance Foundation, has gone through three generations (AutoGen, the OpenAI Agents SDK and a PydanticAI desktop app). Its second version is the most interesting from an architecture standpoint: a lead agent orchestrates data, analysis, modelling, synthesis and report agents plus bull, bear and judge debate agents, but, in the project's words, all numbers come from 32 deterministic Python operators and the LLM only narrates.

That last sentence is the quiet trend of 2026. The first wave let the model do the arithmetic. The current wave increasingly lets the model argue and lets code compute.

Variations on debate

Once the org-chart metaphor existed, the obvious experiment was to vary who is in the room and how they vote.

• Credibility scoring. TrustTrade, a TradingAgents fork from a Harvard lab published in October 2026, inserts a credibility scorer between the analysts and the researchers so that only "selective consensus" reaches the trader. Its paper reports that a high-consensus configuration raised average cumulative return on three large caps in 2024 Q1 from roughly 10% to 26%, at the cost of higher drawdown.
• Councils with vetoes. astra-quant-agent runs a multi-model council (macro, trend, microstructure, contrarian seats) with consensus rules such as "paranoid veto" and "weighted majority", then hands the result to deterministic Python risk gates before anything reaches OKX.
• Scheduled teams. yebof's quant-agent runs eight daily agents (technical, news, macro, earnings, portfolio manager, risk manager, position reviewer, evening analyst) across six market sessions, plus a quarterly "meta reflector" that revises the others.
• Depth levels. TradingAgents-CN exposes five research depths, which is a pragmatic answer to the cost problem: a full debate on every ticker every day is expensive.
• Research loops instead of trading loops. Microsoft's RD-Agent applies two roles, a researcher that proposes hypotheses and a developer that implements them, to Qlib factor and model discovery. It never trades; it tries to make the quant's job faster.

What the architecture is good at

Three things, in our reading of the projects and their papers:

1. Legibility. A debate transcript is a far better artefact than a single model's answer. You can see which evidence the bear used, which the bull ignored, and where the risk team objected. For research, teaching and post-mortems that is valuable regardless of returns.
2. Separation of concerns. Giving each agent a narrow brief (only the 10-K, only the news feed) is a sensible way to manage context windows and to keep the cheap-model, expensive-model split explicit.
3. A natural place for guardrails. The risk-committee stage is where the better projects now put deterministic code: position limits, exposure caps, a last-line veto that no prompt can talk its way past.

What remains unproven

This is the part the star counts do not tell you.

Replicable returns. TradingAgents' README now says that backtest results "are not guaranteed to match any published figure" and asks users to treat the framework "as a research scaffold for studying multi-agent analysis, not as a strategy with a fixed, replicable return." That is the right framing, and it is also an admission that the headline numbers from the paper should not be quoted as if they were a track record.

Live evidence. The one continuous live record we know of in the open, the DXRG paper on two production LLM trading-agent fleets, concluded that "the pre-alpha record did not show a directional edge" and that a fixed 2%/4% bracket beat every discretionary exit the agents produced. Those fleets were single-model systems in crypto, not equity debate pipelines, so the result does not transfer directly; it is simply the only live result of its kind.

Does debate help calibration? Nobody has published a clean ablation we could find in which the same model, on the same data, is run with and without the bull-versus-bear stage, scored on calibrated probability rather than on a backtest return. That is the experiment the field still owes itself, and we will link to the first clean one we find.

Cost. Multi-agent pipelines multiply tokens. TradingAgents' tiering and TradingAgents-CN's depth levels exist because a full debate per ticker per day is not cheap; none of the READMEs we audited states a cost per decision.

Where this is going

Three signals from the current repositories, each of which we will keep tracking on the leaderboard:

• The centre of gravity is moving from "agents decide" to "agents narrate, code decides". FinRobot V2's deterministic operators, astra's physical risk gates and Lumibot's deterministic order gate are all versions of this.
• Memory is becoming a separate layer. FinMem introduced layered memory inside one agent in 2023; TradeMemory Protocol now sells it as an MCP server any agent can plug into, with an audit trail attached.
• The most useful new projects are evaluation harnesses rather than agents: TrustTrade's release scripts, the Kalshi bot's edge harness, DXRG's published fleet data. The field is starting to measure itself, which is the step that comes before it earns the word "investing".

Our own contribution for now is the audit of twenty open-source agents: before asking whether debate helps, it is worth knowing which of these systems can execute at all, under what disclaimers, and with what evidence.