Only 3 of the 20 agentic trading systems hand-coded by this survey make order execution a core function. Signal discovery is core in all 20.
A disclosure comes first. We have no full agent stack with autonomous planning, persistent memory and multi-agent deliberation that we can run against this paper. We can run a controlled factor-search and validation workflow using available price, fundamental and text-embedding features. Any LLM output stops at auditable candidate formulas; autonomous live decisions are outside the workflow. The paper's method therefore remains untested here, and nothing below should be read as such a test.
A map of the research
Hua, Yang and Hao, with authors at HKUST (Guangzhou), Peking University's HSBC Business School and IDEA Research, offer a 9-page map of LLM-agent research in quantitative trading. Agentic has a precise meaning here. The LLM system divides a goal into sub-tasks, calls external tools such as databases and backtest engines, retains task state, and adjusts through feedback.
The proposed source of returns is a chain in which agents generate candidate factors, combine text and price evidence into views, then size and execute them. The five stages describe that chain: factor mining, signal discovery, portfolio construction, order execution, risk management.
The survey has three headline findings. Current systems cluster around signal discovery, while complete integration with portfolio construction, execution and risk control remains, in the abstract's words, "still uncommon". Multi-agent systems depend heavily on aggregation. Strong model or forecasting capability does not reliably translate into trading performance once live market conditions and reliability controls enter. The authors found the 3-of-20 execution count; this review did not uncover it as a fault. Their conclusion asks for what a sceptical reader would want next: connect decisions to execution, then test under conditions that show whether returns are real and reproducible.
The objection survives that concession. Sections 2 and 3 state the coding taxonomies, yet they leave the threshold between core coverage and partial coverage unspecified. The paper also reports no inter-rater agreement. Its directional finding is safe. The exact level, 9 of 20 for risk management and 3 of 20 for execution, remains one team's judgment across 20 papers.
Two pieces of hand coding produce the results. Twenty representative systems, comprising 3 single-agent and 17 multi-agent systems, receive core, partial or no-coverage marks for each stage. They are also classified along three mechanism dimensions: architecture, coordination, adaptation. Fifteen benchmarks are divided among strategy construction, offline trading, live-market evaluation and reliability evaluation.
The authors contribute no dataset of their own, period, universe, cost model or return series. Every quantitative claim is a tally drawn from the paper's two tables, and the corpus is weighted heavily toward 2025 and 2026 preprints. The authors state the boundary directly: these counts cover the systems and benchmarks included in this review.
The arithmetic checks
Rebuilding the tallies from Table 1 takes ten-minute and produces the reported figures exactly. Signal discovery is core in 20 of 20 systems. The corresponding counts are 9 for risk management, 8 for portfolio construction, 4 for factor mining and 3 for order execution.
Four systems cover all five stages at least partially. Two classify all five as core: the orchestration framework built on MCP and A2A protocols, and AlphaCrafter. Within the 17 multi-agent systems, 15 aggregate specialist outputs, 5 use selection, 3 use debate and 2 use an explicit gate. Seventeen of 20 combine more than one architectural structure, while 12 employ an iterative loop.
The arithmetic is clean.
Where does the grading get decided?
Anyone repeating the exercise must make the core-versus-partial distinction without guidance. The choice matters most for risk management, reported at 9 of 20. The paper's risk section groups interventions of very different weight. At one end are agents that estimate a stochastic model and derive a risk measure. At the other are SAE's pre-trade exposure budgets, cooldown periods, slippage limits and venue restrictions. A generous grade pushes risk coverage higher. Requiring a rejected order to be blocked pulls it toward the gating count of 2.
Execution raises the same issue. FinMem, FinAgent and FLAG-Trader each receive partial marks in the execution column, though the paper never defines the boundary between partial execution and core execution. Its one-line descriptions identify layered memory with an adaptive risk persona, dual-level reflection with memory retrieval, and a PPO-tuned LLM policy network.
Reasonable regrading would leave the ordering intact. Signal discovery is unanimous, and execution remains last. The level can move. Beyond the stated inclusion rule, studies where LLM agents contribute to trading workflows, excluding general financial LLM applications and non-agent prediction models, the paper supplies no search string or database.
Aggregation, then the missing machinery
Fifteen of 17 multi-agent systems decide by combining specialist outputs. Debate appears three times. Gating appears twice. An average of analyst opinions yields a sign and a confidence, but a tradeable position still needs the machinery that the coverage counts find scarce: sizing against a risk budget, an order schedule, and a check that blocks trades built from stale inputs.
The authors call for these mechanisms to be compared under the same settings. They conduct no such comparison, and the absence of a shared universe, period or cost model across the 20 systems prevents the survey from doing so. It counts how many papers use each mechanism. Ranking them would require all 20 to run on one universe, one period and one cost model, with realized P&L as the axis. We raised a related complaint about a 30-paper diffusion survey that used four incompatible scorecards (/articles/thirty-diffusion-papers-four-scorecards-nothing-to-compare-them-on).
Nearer the market, thinner evidence
The benchmark taxonomy is the part of the paper I would hand to a colleague. Fifteen benchmarks are classified as 3 strategy construction, 3 offline trading, 6 live-market and 3 reliability. Five of the six live benchmarks report trading outcomes. All five find that realized performance depends on more than model capability alone, including risk control, system design, liquidity and slippage.
Agent Market Arena finds that architecture explains more performance variation than the LLM backbone when inputs are shared and verified. PolyBench finds that accuracy and confidence do not ensure profit, with order-book liquidity and slippage materially changing model rankings. DeepFund reports that most LLMs lost money in live evaluation, where cash management and diversification mattered more than fluent analysis. Among the reliability benchmarks, all three show that leakage, misinformation or attribution changes the meaning of a reported return. KTD-FIN attributes agent returns mainly to market and style exposure rather than stock selection.
The paper never states how long any live benchmark ran or how many decisions each one observed. Five out of five is strikingly consistent, yet the decision counts remain unknown. A live benchmark publishing its decision count and horizon would move me further than another architecture paper.
Our own gap
We have no quote or order-book data. At daily or one-minute bars, spread, market impact, latency and limit-order placement can only be approximated. The survey's thinnest stage is therefore the one stage we cannot test at the fidelity the question demands. This limitation is ours. It repeats the unresolved cost and timing problem we encountered with FinSMART (/articles/where-finsmart-s-returns-come-from).
The paper gives a number to an impression many of us already held. Agentic trading research is largely signal-discovery research inside a portfolio wrapper, while the two systems in the table that carry a decision from factor to fill remain the exceptions.