QuantAgents: Multi-Domain Agentic Systems

• QuantAgents are a family of systems that deploy specialized agents, explicit memory, and tool use to translate high-level objectives into executable quantitative or quantum procedures.
• Their implementations span multi-agent financial trading with simulated trading, decentralized reinforcement learning using quantile coordination, and autonomous quantum programming.
• Innovative techniques such as Time Difference Likelihood in MARL and dual reward mechanisms in finance enable QuantAgents to improve decision-making in complex, domain-specific environments.

QuantAgents is a polysemous research term used for several agentic systems that share a common concern with structured decision-making under domain-specific constraints. In the most direct title usage, it denotes a multi-agent financial system that integrates simulated trading into fund-company-style decision workflows (Li et al., 6 Oct 2025). In a broader technical sense, the term and closely related variants such as QuantAgent have also been used for quantile-based coordination in decentralized reinforcement learning, quantitative-research automation, quantum-security analysis, autonomous machine learning engineering, and quantum-science workflow automation (Lyu et al., 2018, Lim et al., 20 Apr 2026, Alquwayfili, 15 Dec 2025, Sahney et al., 13 Oct 2025, Gustin et al., 21 Dec 2025). Across these usages, the unifying theme is not a single algorithm but a family of architectures in which specialized agents, explicit memory, tool use, and evaluation loops are used to translate high-level objectives into executable quantitative or quantum-domain procedures.

1. Terminological scope and research contexts

The most explicit and title-level use of the plural form appears in “QuantAgents: Towards Multi-agent Financial System via Simulated Trading” (Li et al., 6 Oct 2025). There, QuantAgents is a multi-agent financial trading system over NASDAQ-100 constituents that combines daily stock data, technical indicators, financial news, company reports, and macroeconomic policy information with a workflow centered on simulated trading, risk control, and manager-level synthesis. The same paper frames simulated trading as the critical addition that distinguishes the system from prior LLM-based financial agents focused mainly on “post-reflection” rather than long-term prediction (Li et al., 6 Oct 2025).

A second usage emerges from distributional reinforcement learning. “Likelihood Quantile Networks for Coordinating Multi-Agent Reinforcement Learning” does not use the exact title string QuantAgents, but its detailed discussion explicitly identifies fully decentralized cooperative MARL with quantile value distributions as the setting most relevant to that notion (Lyu et al., 2018). In this line of work, agents coordinate by representing return uncertainty explicitly and by using discrepancies between current and target return distributions as a proxy for teammate-induced non-stationarity.

A third usage expands the term into quantum and scientific-computing contexts. “Adaptive Quantum Matter: Variational Organization through Ising Agents” presents Adaptive Quantum Ising Agents as a minimal Hamiltonian framework for “QuantAgents,” where each agent is a finite transverse-field Ising subsystem with adaptive feedback couplings based on reduced observables (Nagpal et al., 4 Nov 2025). Other papers apply closely related naming to quantum programming, quantum simulation, quantum chemistry, and quantum laboratories, but with different operational meanings: autonomous OpenQASM programming, VQC design, ORCA-based quantum chemistry, self-driving superconducting-qubit calibration, and heterogeneous quantum-simulation orchestration (Fu et al., 26 Aug 2025, Knipfer et al., 22 Feb 2026, Zou et al., 5 May 2025, Cao et al., 2024, Gustin et al., 21 Dec 2025).

This diversity suggests that QuantAgents is best understood as a family resemblance term. In some papers it denotes financial agents concerned with alpha generation, portfolio control, or empirical factor research; in others it denotes quantile-based reinforcement-learning agents; in others it denotes adaptive quantum or quantum-software agents. The common denominator is the delegation of specialized quantitative or quantum reasoning to structured, tool-using, memory-bearing agentic systems.

2. Quantile distributions as a coordination substrate

In decentralized cooperative MARL, the paper “Likelihood Quantile Networks for Coordinating Multi-Agent Reinforcement Learning” studies independent learners in Dec-POMDP-style environments, where each agent receives only its own observations, actions, and team reward, and has no centralized critic or access to teammates’ actions (Lyu et al., 2018). The central difficulty is non-stationarity induced by the exploration and policy change of other agents. Standard independent Q-learning then fits a moving target and becomes vulnerable to shadowed equilibria, alter-exploration, and confusion between genuine environment stochasticity and teammate-induced non-stationarity.

The proposed solution is to replace scalar value estimation with distributional estimation. Rather than estimating only

$Q^\pi(s,a)=\mathbb{E}[Z^\pi(s,a)],$

the agent estimates the return random variable $Z^\pi$ using Implicit Quantile Networks. For sampled quantile fraction $\tau\sim U([0,1])$, the network predicts

$F^{-1}_{\pi,\tau}(s,a),$

and expected action values are recovered through distorted or undistorted averaging over quantiles. Learning uses the quantile Huber loss

$\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$

with pairwise TD errors between online and target quantile samples (Lyu et al., 2018).

The distinctive coordination mechanism is Time Difference Likelihood. From online quantile samples $d_{1:M}$ and target samples $t_{1:M'}$, the paper constructs a piecewise linear approximation to the quantile CDF and defines a sample-set likelihood measuring how much probability mass of the current estimated return distribution lies near the target return samples. This TDL score is then used to modulate negative updates through a likelihood-hysteretic learning rule:

$\mu_t= \begin{dcases} \max(\beta,l_{t_{1:M'},d_{1:M})\,\bar{\mu}, & \text{if } \delta^{\tau,\tau'}_t \le 0 \ \bar{\mu}, & \text{otherwise} \end{dcases}$

so that unlikely negative samples exert less downward pressure on values (Lyu et al., 2018).

The same framework separates optimism in value learning from optimism in policy derivation through Dynamic Risk Distortion. The paper uses two operators, Conditional Value-not-at-Risk

$\mathrm{CVnaR}(\eta,\tau)=1-\eta\tau$

and Wang distortion

$$\mathrm{Wang}(\eta,\tau)=\Phi(\Phi^{-1}(\tau)+\eta),$$

with annealing toward risk neutrality over training. TDL/hysteresis changes how negative evidence is absorbed, whereas distortion changes which quantiles are emphasized for action selection (Lyu et al., 2018).

Empirically, the paper reports improved training stability, better sample efficiency, and a higher probability of converging to the joint optimal policy than HDRQN and LDQN on meeting-in-a-grid, CMOTP, and higher-dimensional graphical variants (Lyu et al., 2018). It also reports a particularly important ablation: TDL is much higher when the current agent itself is exploring than when teammates are exploring while the current agent is not, suggesting that local return distributions can discriminate local mistakes from teammate-induced non-stationarity. The authors’ own summary is that quantile return distributions become a coordination mechanism rather than merely a better critic.

3. QuantAgents in quantitative finance and trading

In the financial literature, QuantAgents most directly denotes a fund-company-style multi-agent trading architecture with four specialized roles: Bob as simulated trading analyst, Dave as risk control analyst, Emily as market news analyst, and Otto as manager (Li et al., 6 Oct 2025). These agents interact through three meeting mechanisms: the Market Analysis Meeting, the Strategy Development Meeting, and the Risk Alert Meeting. The system uses 26 tools, 3 types of memory, and 10 actions, and integrates simulated trading directly into policy improvement rather than treating it as an offline benchmark (Li et al., 6 Oct 2025).

The formal core of that system is a dual reward mechanism combining real-market and simulated-trading feedback:

$Z^\pi$0

with adaptive weights

$Z^\pi$1

Risk escalation is handled by a separate portfolio-risk score,

$Z^\pi$2

which triggers the Risk Alert Meeting when $Z^\pi$3 (Li et al., 6 Oct 2025). On the 2021–2023 NASDAQ-100 test period, the paper reports ARR $Z^\pi$4, TR $Z^\pi$5, SR $Z^\pi$6, CR $Z^\pi$7, SoR $Z^\pi$8, MDD $Z^\pi$9, Vol $\tau\sim U([0,1])$0, ENT $\tau\sim U([0,1])$1, and ENB $\tau\sim U([0,1])$2 (Li et al., 6 Oct 2025).

A separate line uses the singular QuantAgent for self-improving alpha mining. “QuantAgent: Seeking Holy Grail in Trading by Self-Improving LLM” introduces a two-layer framework in which an inner loop performs KB-augmented writer–judge refinement and an outer loop evaluates generated signals in a real environment and updates the knowledge base (Wang et al., 2024). The paper formalizes the inner loop as an MDP with value functions

$\tau\sim U([0,1])$3

and states a sublinear Bayesian-regret claim for the overall framework under strong assumptions imported from prior work (Wang et al., 2024). In the trading instantiation, the system operates on 500 Chinese A-share stocks in 2023, uses gpt-4-0125-preview, generates factor code under a predefined Factor template, and evaluates signals using metrics such as IC and Sharpe (Wang et al., 2024).

QRAFTI shifts the same financial-agentic idea away from live trading toward empirical factor research. It is a multi-agent framework for panel-data-based equity factor research, with a Factor Research Agent, a Standardized Reporting Agent, and a Code Writing Agent, plus reflection-based planning and MCP-exposed financial tools (Lim et al., 20 Apr 2026). For multi-step tasks, the paper reports Sim@1 $\tau\sim U([0,1])$4 with all tools, improving to $\tau\sim U([0,1])$5 with all tools plus reflection planning, and coding-only configurations perform materially worse on 1-step, 2-step, and multi-step tasks (Lim et al., 20 Apr 2026). This suggests that in quantitative finance, agentic systems are being used not only for portfolio decisions but also for reproducible research, factor replication, and standardized anomaly evaluation.

A more execution-oriented short-horizon variant appears in “QuantAgent: Price-Driven Multi-Agent LLMs for High-Frequency Trading,” which uses IndicatorAgent, PatternAgent, TrendAgent, and a final DecisionAgent over 4-hour OHLC data and a 3-candlestick prediction horizon (Xiong et al., 12 Sep 2025). Its technical-analysis core includes EMA, MACD, RSI, ROC, Stochastic Oscillator, Williams %R, support and resistance lines, and a fixed stop-loss width $\tau\sim U([0,1])$6, with risk–reward target $\tau\sim U([0,1])$7, $\tau\sim U([0,1])$8 (Xiong et al., 12 Sep 2025). The paper reports accuracy improvements over a random baseline on eight assets, with especially large gains on SPX, QQQ, ES, and NQ, but also explicitly notes that the setup uses 4-hour bars and is therefore better read as short-horizon algorithmic trading than true microstructure-level HFT (Xiong et al., 12 Sep 2025).

4. Quantum-software, quantum-simulation, and quantum-chemistry QuantAgents

In quantum programming and simulation, closely related agentic systems use the same pattern of specialist roles, documentation grounding, tool calling, and iterative repair. “QAgent: An LLM-based Multi-Agent System for Autonomous OpenQASM programming” uses a hierarchical framework with a Dynamic-few-shot Coder and a Tools-augmented Coder, coupled by a fallback policy (Fu et al., 26 Aug 2025). The system integrates task planning, RAG, in-context few-shot learning, predefined generation tools, test generation, CoT-style reflection, and iterative repair, and the paper summarizes its performance as up to a 71.6% improvement over previous static LLM-based approaches (Fu et al., 26 Aug 2025).

“AI Agents for Variational Quantum Circuit Design” applies a single tool-using LLM agent to VQC architecture search rather than quantum-code synthesis (Knipfer et al., 22 Feb 2026). The agent proposes candidate PennyLane circuits, triggers automated training and validation, receives structured metrics such as RMSE, gate count, and parameter count, and updates its design strategy from performance feedback. The reported behavior is not monotonic optimization but gradual evolution from simple ansätze to more expressive structures, with motifs such as data/computation qubit separation, star-topology entanglement, selective measurement, and recurrent-style data re-uploading emerging from iterative search (Knipfer et al., 22 Feb 2026).

At the workflow level, “El Agente Q: An Autonomous Agent for Quantum Chemistry” and its later extension “El Agente Quntur: A research collaborator agent for quantum chemistry” show the same QuantAgents pattern in computational chemistry (Zou et al., 5 May 2025, Pérez-Sánchez et al., 4 Feb 2026). El Agente Q uses a hierarchical memory framework with a computational_chemist top-level agent, geometry, quantum-calculation, and file-I/O modules, ORCA-specific block experts, shell access, SLURM submission, and in situ debugging. On six university-level benchmark exercises, it reports $\tau\sim U([0,1])$9 average success on first-level tasks and $F^{-1}_{\pi,\tau}(s,a),$0 on second-level tasks, together with concrete recovery from hallucinated ORCA keywords, wrong XYZ references, and imaginary frequencies (Zou et al., 5 May 2025). Quntur expands this into a 58-agent, 34-tool system around ORCA 6.0, guided deep research, and literature reasoning, and presents it explicitly as a research collaborator rather than merely an automation tool (Pérez-Sánchez et al., 4 Feb 2026).

The same pattern extends to real or simulated quantum experiments. “Agents for self-driving laboratories applied to quantum computing” introduces k-agents, which combine knowledge agents with an execution agent that decomposes procedures into state machines and closes the loop through visual and text inspection of experiment results (Cao et al., 2024). On a superconducting processor, the agents autonomously calibrated and operated hardware, searched siZZle-gate parameters for 3 hours and up to 100 experiments, identified working settings at 4726 MHz and amplitude 0.3049, and produced and characterized a three-qubit GHZ state (Cao et al., 2024). “El Agente Cuántico: Automating quantum simulations” generalizes this style to heterogeneous software frameworks—CUDA-Q, PennyLane, Qiskit, QuTiP, TeNPy, and Tequila—behind a single natural-language interface, with a quantum scientist orchestrator delegating to package-specific experts and using literature search, Deepsearch over documentation, Python execution, and HPC submission (Gustin et al., 21 Dec 2025).

A different but related line uses agents as a front end for explainable surrogate modeling. xChemAgents couples a Selector agent and a Validator agent to a multimodal molecular predictor, forcing sparse descriptor choice $F^{-1}_{\pi,\tau}(s,a),$1, simplex weights $F^{-1}_{\pi,\tau}(s,a),$2, and a physics-aware text embedding

$F^{-1}_{\pi,\tau}(s,a),$3

before fusion with an equivariant geometric encoder (Polat et al., 26 May 2025). The paper reports up to a 22% reduction in MAE over strong baselines on selected QM9 properties, especially HOMO/LUMO/gap-related targets, while attaching natural-language rationales and physical checks to descriptor selection (Polat et al., 26 May 2025).

5. Shared architectural themes and systems primitives

Across domains, several architectural regularities recur. One is role specialization with explicit orchestration. Quantigence, a multi-agent AI framework for quantum-security research, defines a Supervisor, Cryptographic Analyst, Threat Modeler, Standards Specialist, and Risk Assessor, coordinated by a LangGraph finite-state machine with dependency-aware task decomposition, critique-and-retry loops, and a shared memory graph (Alquwayfili, 15 Dec 2025). Its main formal artifact is the Quantum-Adjusted Risk Score, extending Mosca’s theorem via

$F^{-1}_{\pi,\tau}(s,a),$4

a temporal urgency factor

$F^{-1}_{\pi,\tau}(s,a),$5

and a weighted score

$F^{-1}_{\pi,\tau}(s,a),$6

with default weights $F^{-1}_{\pi,\tau}(s,a),$7, $F^{-1}_{\pi,\tau}(s,a),$8, $F^{-1}_{\pi,\tau}(s,a),$9 (Alquwayfili, 15 Dec 2025). The same paper operationalizes “cognitive parallelism” as role-isolated reasoning under serialized execution on a single RTX 2060 SUPER, using 4-bit Qwen2-7B-Instruct and context swapping rather than real concurrent inference (Alquwayfili, 15 Dec 2025).

A second recurring theme is that some supposedly multi-stage workloads collapse effectively into a strong single-agent design when context continuity dominates specialization. Operand Quant is a single-agent IDE-based architecture for autonomous machine learning engineering, operating through a structured JSON action loop, asynchronous process monitoring, persistent IDE state, and context compaction (Sahney et al., 13 Oct 2025). On MLE-Benchmark (2025) it reports an overall medal rate of $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$0 across 75 problems, with subset rates $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$1, $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$2, and $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$3 for Lite, Medium, and Hard, respectively (Sahney et al., 13 Oct 2025). This suggests that the QuantAgents design space includes both explicit societies of agents and carefully instrumented single agents with unified reasoning state.

A third theme is that agent performance increasingly depends on inference-time systems engineering. UltraQuant studies 4-bit KV-cache compression for context-heavy, multi-round agent workloads and reports that on a long-context, multi-turn agentic workload, UltraQuant cuts P50 time-to-first-token by $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$4 in late rounds, $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$5 across all rounds, and raises output throughput by $\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$6 over an FP8 KV baseline (Chakrabarti et al., 18 Jun 2026). The paper’s point is explicitly agent-centric: in multi-round systems, dynamic memory state rather than static model weights becomes a first-order serving bottleneck.

A fourth theme is that uncertainty is being moved from passive diagnosis into active control. “Agentic Uncertainty Quantification” defines a Dual-Process Agentic UQ framework with System 1 Uncertainty-Aware Memory

$\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$7

and System 2 Uncertainty-Aware Reflection triggered by

$\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$8

with trajectory-level calibration metrics such as T-ECE and T-BS (Zhang et al., 22 Jan 2026). On ALFWorld and WebShop, the full AUQ system outperforms ReAct, Reflexion, Self-Reflection, and CoT-SC in success rate and trajectory-level calibration (Zhang et al., 22 Jan 2026). This suggests that reliable QuantAgents increasingly treat confidence not as a report to the user but as a routing variable that governs memory use, retrieval depth, and reflective compute.

A fifth theme is the widening of the notion of “agent” itself. AQIA models each agent as a finite TFIM patch with summary vector

$\rho_{\tau}(\delta) = (\tau-\mathds{1}_{\delta\le 0})\frac{\mathcal{H}_\kappa(\delta)}{\kappa},$9

adaptive similarity-based couplings, and a self-consistency map $d_{1:M}$0, yielding domain formation, adaptive glass, and modular polarization without fixed geometric embedding (Nagpal et al., 4 Nov 2025). The paper does not demonstrate task learning in the strong machine-learning sense, but it does formalize memory-like hysteresis, feedback-driven adaptation, and collective organization. A plausible implication is that some QuantAgents research is shifting from software agents toward physically grounded adaptive subsystems.

6. Limitations, controversies, and open directions

The most immediate limitation is that the term itself is not stable. It denotes distributional MARL in one context, simulated-trading agents in another, short-horizon technical trading in another, panel-data factor research in another, and quantum-software or quantum-matter systems elsewhere (Lyu et al., 2018, Li et al., 6 Oct 2025, Xiong et al., 12 Sep 2025, Lim et al., 20 Apr 2026, Nagpal et al., 4 Nov 2025). This semantic dispersion is not merely terminological; it reflects distinct notions of agency, memory, control, and evaluation.

Within each line of work, the limitations are substantial. In quantile-based MARL, TDL depends on immature quantile estimates early in training, uses a heuristic piecewise-linear likelihood approximation, assigns zero density outside the sampled range, and comes without a theoretical convergence proof (Lyu et al., 2018). In the financial QuantAgents paper, the authors explicitly acknowledge underexplored generalization and possible LLM information leakage in backtesting, while exact functional forms for several reward components and portfolio-weight construction are not specified (Li et al., 6 Oct 2025). In the self-improving trading QuantAgent paper, the outer loop is exposed to obvious data-snooping and regime-specific overfitting risks because the empirical protocol does not provide a stringent out-of-sample trading evaluation (Wang et al., 2024). In the short-horizon QuantAgent paper, the title language of HFT is weakened by a concrete setup based on 4-hour bars, no order-book inputs, no transaction-cost model, and only a random baseline (Xiong et al., 12 Sep 2025).

In the quantum-science line, the constraints are different but equally serious. QAgent’s pass@k formula is malformed in the manuscript, decomposition degrades sharply on level-3 and level-4 composite OpenQASM tasks, and the precise basis of the 71.6% improvement is not numerically decomposed in the main text (Fu et al., 26 Aug 2025). VQC-design agents remain simulation-bound, susceptible to context-window failures, and far from real noisy-hardware closure (Knipfer et al., 22 Feb 2026). El Agente Q and Quntur still struggle most with geometry-heavy transition-state and path-generation tasks, and the authors explicitly note overconfident post-analysis and incomplete safeguards in mathematical post-processing (Zou et al., 5 May 2025, Pérez-Sánchez et al., 4 Feb 2026). El Agente Cuántico presents broad workflow coverage, but its evidence is still a collection of representative sessions rather than a large, standardized benchmark across all claimed simulation paradigms (Gustin et al., 21 Dec 2025).

Architectural and evaluation issues also recur. Quantigence reports a 67% reduction in research turnaround time and higher source coverage, but does not provide detailed timing methodology, formal agreement analysis, or ablations isolating supervisor review and retrieval (Alquwayfili, 15 Dec 2025). QRAFTI shows strong gains from reflection planning and tool grounding, yet evaluates only a small number of empirical-finance tasks in depth and leaves sample ranges and universes under-specified (Lim et al., 20 Apr 2026). xChemAgents claims faithful explanations, but does not provide formal faithfulness tests, deletion analyses, or counterfactual auditing of its Selector–Validator rationales (Polat et al., 26 May 2025). AUQ improves calibration and success rates, but the paper explicitly identifies “Delusional Confirmation” as a failure mode in which System 2 reflection can increase confidence around a wrong plan (Zhang et al., 22 Jan 2026).

These limits point toward a shared research agenda. QuantAgents systems increasingly require stronger validation protocols, explicit handling of distribution shift, better uncertainty gating, more faithful explanation auditing, and tighter coupling between reasoning traces and executable verification. This suggests that the field is moving from simple role decomposition toward a harder problem: how to make domain-specialized agents not only act, but remain scientifically, financially, or operationally trustworthy over long horizons.