Re-deriving Agent Roles When a Trading System Changes Markets

Teams of language-model agents that trade, with one agent mining signals, another judging the market regime, and another sizing positions, are usually designed for one market and then carried unchanged into others. We think the part that fails to carry over is often the division of labor itself, not the model weights or the prompts. A Chinese A-share desk has to manage work that a U.S. desk never does. Shares bought today cannot be sold until tomorrow, a stock pinned at its ±10% daily limit builds a queue of unfilled orders, and short selling is restricted. A crypto desk faces 24/7 trading and thin weekend books. Vocational education research has a name for this mismatch: training that is organized by topic instead of by the actual operational process of the job produces graduates who know the material but cannot do the work. It also has a concrete procedure for fixing it, which derives roles and step sequences from the target workplace's real job functions.

We propose to bring that procedure to multi-agent trading. Before deploying in a new market, the system reads the market's rulebook and microstructure. It writes out the work a desk in that market must perform each decision cycle, then rebuilds the agent graph from that process map instead of reusing the source market's graph. The claim is narrow so that it can be tested. Injecting market rules into existing agents' prompts should be enough when a rule is only a constraint that filters actions. Re-deriving roles should matter when a rule creates new dependent work, such as the planning that T+1 settlement forces on inventory, or estimating queue position at a price limit. A simulator in which we can switch individual rules on and off lets us test that split directly.

Where This Sits

Three financial multi-agent systems motivate the problem. AlphaCrafter chains a factor-mining agent, a regime-aware screener and a trader, and evaluates them on both CSI 300 and S&P 500. QuantAgents runs four analyst and manager agents through structured meetings, with backtests on NASDAQ-100 and live trading in A-shares and Hong Kong stocks. Both use one hand-built role graph in every market, and QuantAgents' authors say generalization to unseen regimes needs further work. The strongest motivating evidence comes from When Agents Trade, a live benchmark on two U.S. stocks and two crypto tokens. It found that changing the agent architecture shifted behavior and returns far more than swapping the underlying language model. If architecture dominates, then the architecture's fit to the market is the obvious thing to adapt, yet none of these systems adapts it.

HiveMind is the closest adaptive method, and it supplies a tool we reuse. It treats the agent team as a directed acyclic graph, which is a fixed one-way flow of messages between agents. It scores each agent's marginal contribution with a pruned Shapley computation (DAG-Shapley), and then rewrites the prompts of weak agents. HiveMind adapts prompts on a fixed topology, and only within U.S. mega-cap tech. Its prompt updates also raised volatility and drawdowns, so it acts as a constraint on our design: any adaptation we test must run under the same explicit risk budget. Agent Trading Arena and DeMAgent provide the environment. Both use simulated exchanges with endogenous price formation and evaluate on NASDAQ, CSI and crypto. They also show that chart inputs and code execution sidestep numerical-reasoning failures, so we keep both capabilities fixed across all conditions. Both use essentially a single agent with no decomposed roles.

The vocational literature supplies the method. In Wang et al., a repair course is rebuilt in three stages: actual job functions from workplace surveys (the action field), then educational objectives, then concrete classroom scenarios, all run through a six-step loop of message, decision, plan, implement, check and evaluate. He et al. map student groups onto real workplace roles on a real building project. JCRec decomposes job postings into skills with a language model. It reports only 0.68 accuracy at extracting proficiency levels, which warns us that automated process extraction must be audited by humans. The vocational studies themselves lack quantitative evidence that graduates need less on-the-job training. Our setting can measure the agent analogue of that quantity, which is how much target-market experience a system needs before it performs competently.

We searched for prior work joining all three directions and found none. The novelty here is in what gets adapted on transfer, namely the role graph derived from the target market's work process. Adding agents or adding markets would not be new.

Core Claim

When a multi-agent trading system moves to a market whose rules create new dependent work steps, rebuilding the role graph from a written process map of that market will beat three alternatives under matched agent count and call budget. The alternatives are porting the source graph, porting it with the rules injected into prompts, and porting it with HiveMind-style prompt optimization. For rules that only constrain actions, the advantage should vanish. A graph derived from the wrong market's process should do no better than the ported graph. That last prediction separates "the decomposition matches the market" from "any re-decomposition helps."

Research Program

The first stage is a diagnostic. We take a fixed source-market team in the style of AlphaCrafter, built on U.S. equities, and run it unchanged in A-share-like and crypto-like conditions. We compute DAG-Shapley contributions per agent in each market. If decomposition mismatch is real, we expect a specific signature on transfer: some roles become near-zero contributors, and the remaining roles show more rule violations and failed orders. A uniform drop across all agents would not fit the mismatch story. This stage costs little and can kill the project early.

The second stage builds the derivation procedure. For each market, the inputs are a structured rulebook covering trading hours, settlement cycle, price limits, short-sale and borrow rules, lot sizes and fee schedule, plus market-specific desk job descriptions where we can collect them. A language model drafts the action field, meaning the list of tasks a desk must do each cycle with their dependencies, and a person with trading experience audits the draft. Tasks are grouped into roles by shared inputs and tools. Roles are ordered along the six-step loop. The "check" step becomes an explicit agent that verifies each order against that market's rules before execution. The output is a role graph plus role prompts produced by a fixed template, so the human contribution is confined to the audited process map. We release those maps.

The third stage creates counterfactual markets. We extend the endogenous-price simulator from the Agent Trading Arena suite so that individual rules can be toggled while the population of simulated participants stays fixed: daily price limits, T+1 settlement, short-sale bans, 24/7 sessions and borrow fees. Before running anything, we label each rule as constraint-only (a short ban mostly removes actions) or process-creating (T+1 and limit-up queues add planning and estimation work). Because the participants are identical, any difference between graphs is attributable to rules rather than to differences in historical data. The simulator also avoids the pretraining look-ahead problem that QuantAgents acknowledges for its historical backtests.

The fourth stage validates on real data. We use PRUDEX-Compass PRUDEX-Compass for its four-market coverage of U.S. stocks, Chinese stocks, crypto and FX, and its six evaluation axes. All test windows fall after the backbone model's training cutoff. Any news or macro inputs are restricted to decision-time information, following the leakage-aware forecasting benchmark, which shows retrieval-augmented models confusing forecasting skill with look-ahead.

Baselines

The naive reference point is the source-market graph ported unchanged with its original prompts. Every other condition is reported relative to it.

The strongest simple competitor is the ported graph with the target market's full rulebook injected into every agent's prompt, together with a generic rule-check filter that rejects illegal orders. It is cheap and transparent, and for constraint-only rules it may well be optimal. It receives the same rulebook text our derivation uses, the same backbone, the same tools and the same call budget per decision.

The strongest existing adaptive method is HiveMind's contribution-guided prompt optimization on the ported graph, run in the target market. We give it the same number of target-market adaptation episodes our method uses for calibration, and both run under one shared risk-budget rescaling of positions so that HiveMind's known drawdown inflation does not decide the outcome.

We also re-implement fixed designs in the style of AlphaCrafter and QuantAgents as off-the-shelf architectures. For reference, QuantAgents reports a Sharpe ratio of 3.11, an annualized return of 58.68% and a maximum drawdown of 16.86% in its NASDAQ-100 PRUDEX backtest. ArenaTrader reports Sharpe ratios of 0.348 on NASDAQ and 0.123 on CSI in Agent Trading Arena. We found no reported result on Lacuna for AlphaCrafter's specific metrics in the file. These numbers come from different windows and cost assumptions, so we rerun every system in our setting rather than compare to published figures.

Two ablations isolate the key ingredient. In the first, a same-size generic graph has the same number of agents and the same call budget but roles that do not come from any process analysis. This control rules out "more agents helps." In the second, a mismatched process graph applies, for example, the crypto-derived graph to the A-share condition. This control rules out "any re-decomposition helps." Our method beats the field only if all of the following hold on process-creating rules, as paired differences over seeds and windows with bootstrap confidence intervals excluding zero: its cost-adjusted Sharpe exceeds that of the rule-injected port, its rule-violation rate is lower, and the mismatched graph fails to match it.

Why not just…?

Why not just put the market rules in the prompt? This is the real threat, and the experiment is designed around it. For constraint-only rules we expect rule injection to tie our method, and we will report that result. The open question is whether a rule that adds work, such as tracking which shares are locked until tomorrow or estimating whether a limit-down sell order will fill, can be absorbed by agents whose roles were never assigned that work. The multi-agent survey documents context overload and tool confusion when single agents carry too many responsibilities, which suggests absorption will degrade. The constraint-only versus process-creating split in the simulator settles this. If injection ties our method on process-creating rules too, the core claim is false.

Why not just let HiveMind-style prompt optimization adapt the ported team? Prompt optimization can repair a weak agent, but it cannot create a role that the graph lacks, and in its own experiments it increased drawdowns. We give it equal adaptation budget and an equal risk budget. If it closes the gap, then topology matters less than we think. The DAG-Shapley diagnostic will show whether it closes the gap by quietly turning an existing agent into the missing role.

Why not just use one strong model with all the information? The live benchmark found architecture outweighing backbone choice, and Agent Trading Arena found text-only models failing at numeric reasoning. Even so, we include a single-agent baseline with the full rulebook, chart inputs and code execution at a matched call budget, because recent long-context models may narrow this gap.

Why not just retrain a specialized model or reinforcement-learning policy per market? Per-market retraining of a model like WaveLSFormer or a PPO agent does adapt to the new market, and we include a PPO baseline in the simulator. But PRUDEX-Compass reports large drawdowns and regime collapse for standard reinforcement-learning traders, and DeMAgent's code-generating agent outperformed PPO and TD3 across its markets. These baselines also need far more target-market data, which is exactly the adaptation cost we measure.

Proposed Experiments and Decisive Checks

The main experiment covers four simulated rule regimes: U.S.-like, A-share-like with limits plus T+1 plus a short ban, crypto-like with 24/7 sessions and borrow fees, and single-rule toggles. Each pair is run in both transfer directions, with five or more seeds, all conditions sharing one backbone, and a nonzero cost and slippage model. Primary metrics are cost-adjusted Sharpe ratio and maximum drawdown. Secondary metrics are the rule-violation rate, the PRUDEX reliability axis (variance across seeds), and the number of adaptation episodes or language-model calls needed to reach 90% of the target market's eventual performance. That last metric is our measurable stand-in for on-the-job training time.

The decisive check is an interaction effect. Our method's advantage over the rule-injected port should be significantly larger on process-creating rules than on constraint-only rules, and the mismatched-graph ablation should show no advantage. If the advantage appears uniformly, the explanation is something other than process fit, most likely added structure. If it never appears, decomposition is not the transfer bottleneck. We then replicate the ordering on post-cutoff windows of PRUDEX-Compass's U.S., Chinese and crypto data, accepting smaller effects but requiring the same sign.

A final robustness check replaces the human-audited process maps with unaudited model-drafted ones. JCRec's 0.68 extraction accuracy suggests these will be noisier, and this check tells us how much of any gain depends on expert input.