InvestAgent: Autonomous Investment Systems

• InvestAgent is a family of agentic investment systems that combine LLMs, quantitative models, retrieval pipelines, and governance layers to support autonomous research and market execution.
• These systems employ specialized architectures that separate research and decision-making from execution, ensuring auditability, risk control, and compliance.
• Empirical studies report strong performance metrics, though challenges like oracle dependence and scalability bottlenecks remain critical for real-world deployment.

InvestAgent is a label used in recent arXiv literature for several agentic investment systems that combine LLMs, quantitative models, retrieval pipelines, optimization routines, and execution or compliance layers to support investment research and action. Depending on the paper, the term denotes an autonomous investment agent in tokenized alternative asset markets, an LLM aligned with investor decision-making under herd behavior, a solver-centric portfolio optimization system, a multi-agent trading framework with simulated trading, or a due-diligence and screening platform for portfolio formation (Borjigin et al., 30 Jun 2025, Wang et al., 9 Jul 2025, Dangi, 2013, Li et al., 6 Oct 2025, Caner et al., 24 Mar 2026, Alexandrou et al., 13 May 2026). In the literature surveyed here, this suggests that InvestAgent is not a single standardized architecture but a family of agentic investment designs spanning screening, allocation, execution, monitoring, and governance.

1. Terminology and conceptual scope

In one line of work, InvestAgent is defined as an autonomous investment agent built atop an AI-governed, multi-agent tokenization framework for alternative assets. In that formulation, it consumes verification, valuation, compliance, and monitoring outputs; executes compliant on-chain trades; and remains subordinated to a governance layer that can pause, slash, or re-verify when anomalies or policy breaches arise (Borjigin et al., 30 Jun 2025). In another line of work, InvestAgent is an LLM fine-tuned by InvestAlign to align with investor decision-making processes under herd behavior, producing time-indexed portfolio decisions from risk aversion, herding intensity, and market parameters (Wang et al., 9 Jul 2025). A third usage derives from computationally guided portfolio optimization, where InvestAgent denotes an agent-based system integrating solver agents, a solution bank, and a Simulated Annealing super-agent for portfolio construction (Dangi, 2013).

The same label is also applied to role-based multi-agent systems for simulated trading, stock analysis, portfolio screening, venture due diligence, and financial-agent orchestration. These implementations include four-agent systems with simulated trading and risk meetings, multimodal market simulators with execution and bankruptcy checks, two-agent screening systems fused with high-dimensional precision matrix estimation, event-driven due-diligence pipelines on n8n, and planner–orchestrator–alpha–risk–portfolio–execution stacks for algorithmic trading (Li et al., 6 Oct 2025, Zhang et al., 2024, Caner et al., 24 Mar 2026, Alexandrou et al., 13 May 2026, Li et al., 1 Dec 2025). The literature therefore treats InvestAgent less as a fixed product category than as a reusable pattern for organizing specialized financial agents around a common investment objective.

2. Architectural patterns

Despite domain differences, the architectures recur around functional specialization, explicit coordination, and auditable state transitions. The tokenized-asset version includes an Asset Owner Agent, Verification Agent, Valuation Agent, Compliance Agent, Tokenization Agent, Monitoring Agent, and AI Governance Agent, with provenance hashes and approvals anchored on-chain and enforcement delegated to a Governance Smart Contract (Borjigin et al., 30 Jun 2025). QuantAgents organizes a Manager, a Simulated Trading Analyst, a Risk Control Analyst, and a Market News Analyst, coordinated through a weekly Market Analysis Meeting, a weekly Strategy Development Meeting, and an event-driven Risk Alert Meeting (Li et al., 6 Oct 2025). The orchestration framework for financial agents expands this decomposition to planner, orchestrator, alpha agents, risk agents, portfolio agents, backtest agents, execution agents, audit agents, and a memory agent, with MCP and A2A message passing (Li et al., 1 Dec 2025).

Other variants emphasize reflective or deterministic reasoning rather than governance contracts. FinAgent is organized around Market Intelligence, Memory, Low-level Reflection, High-level Reflection, and Tool-augmented Decision-making, with multimodal inputs and diversified memory retrieval (Zhang et al., 2024). AlphaAgents uses a Coordinator and a round-robin debate among Fundamental, Sentiment, and Valuation agents, requiring each agent to speak at least twice before termination (Zhao et al., 15 Aug 2025). GuruAgents uses a deterministic reasoning pipeline in which role-based prompts emulate Warren Buffett, Benjamin Graham, Joel Greenblatt, Joseph Piotroski, or Edward Altman, forcing tool use before ranking and constraining output to a standardized portfolio table (Kim et al., 2 Oct 2025).

A second common pattern is explicit separation between analysis and execution. Screening and due-diligence systems often stop at a narrowed candidate set or a structured research report, whereas trading systems extend to position adjustment, order placement, and post-trade monitoring. The three-layer screening architecture with an LLM-S fundamentals agent, a FinBERT sentiment agent, and a precision-matrix weighting engine exemplifies this separation, as does the venture-capital due-diligence pipeline that terminates in HTML reports and analyst assessments rather than direct market execution (Caner et al., 24 Mar 2026, Alexandrou et al., 13 May 2026). This suggests that “InvestAgent” spans both executable trading agents and upstream research agents whose output is investable intelligence rather than orders.

3. Learning, retrieval, and signal formation

One technically explicit formulation of InvestAgent is the herd-behavior model in InvestAlign. There, the market and budget dynamics are

$$dX_i(t)=\big[rX_i(t)+vP_i(t)\big]dt+\sigma P_i(t)\,dW(t),$$

with CARA utility

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$

and herding penalties

$$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$

For the simple problem $P3$, the theoretical supervision targets are

$$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$

$$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$

with $\eta>0$ computed numerically. These theoretical solutions are transformed into textual SFT pairs for LLM fine-tuning, and the paper argues that noiseless theoretical targets produce steeper early gradients than real-user targets (Wang et al., 9 Jul 2025).

A different research line uses retrieval and multimodal reflection rather than analytic solutions. FinAgent treats trading as an MDP but augments the policy with internal reasoning modules:

$$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$

while its Market Intelligence, Low-level Reflection, and High-level Reflection modules consume prices, news, expert guidance, Kline charts, and trading charts (Zhang et al., 2024). FinArena likewise adopts a mixture-of-experts-inspired human–agent framework in which a Time Series Agent, a News Agent with adaptive RAG, and a Statement Agent feed a universal expert that conditions decisions on a user risk vector $\mathbf{R}$ (Xu et al., 4 Mar 2025).

Retrieval grounding is also central in enterprise and due-diligence settings. AI PB combines OpenSearch with a finance-domain embedding model, deterministic routing between internal and external LLMs, and post-generation evidence validation requiring reference tokens for statements; the system reports more than 30% hallucination reduction versus vanilla prompting in internal tests (Park et al., 23 Oct 2025). The venture-capital due-diligence framework uses Perplexity Sonar for live web intelligence, programmatic retrieval from $\Gamma$.E.MI. dynamic endpoints, and layout-aware OCR with Marker, while explicitly marking “Not Found” when official or third-party financials are absent rather than synthesizing unverifiable figures (Alexandrou et al., 13 May 2026). In tokenized alternative assets, verification relies on secure oracles, authenticated data feeds, document hashes, registry references, and cross-verification by multiple independent agents (Borjigin et al., 30 Jun 2025).

Signal formation can also be constrained by financial priors rather than free-form reasoning. AlphaAgent imposes originality through AST similarity, semantic alignment between hypotheses and generated factors, and complexity penalties:

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$0

with

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$1

This is a distinct interpretation of InvestAgent: not an execution framework, but an autonomous alpha-mining engine whose product is decay-resistant predictive factors rather than orders (Tang et al., 24 Feb 2025).

4. Portfolio construction, execution, and risk control

Portfolio construction mechanisms vary sharply across InvestAgent variants. The computationally guided portfolio system centers on canonical optimization programs such as mean–variance,

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$2

as well as Sharpe maximization, CVaR minimization, equally weighted, global minimum variance, most diversified portfolio, and equally weighted risk contribution. Its super-agent uses an enhanced Simulated Annealing search guided by a solution bank and a decision matrix encoding heuristic knowledge (Dangi, 2013). The screening architecture for S&P 500 portfolios instead treats the screened universe size $$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$3 as a random variable, then applies high-dimensional precision matrix estimation to build GMV, MV, or MSR portfolios on the screened set (Caner et al., 24 Mar 2026).

Several systems translate predictions into discrete position ladders rather than continuous optimization. In the Chinese Public REITs framework, the Prediction Agent outputs $$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$4 for $$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$5, and the Decision Agent maps a horizon-weighted score

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$6

into stepwise exposure changes

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$7

with transaction cost $$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$8 per trade (Li, 22 Jan 2026). QuantAgents uses a different risk gate: a Risk Alert Meeting is triggered when

$$\phi_i[X_i(T)]=-\frac{1}{\alpha_i}\exp\big[-\alpha_iX_i(T)\big],$$9

after which the policy is updated with a risk-aware reward term (Li et al., 6 Oct 2025).

Execution environments are equally heterogeneous. StockAgent simulates exchange sessions, order-book matching, random clock scheduling, transaction costs of $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$0 per share with minimum $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$1 and maximum $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$2 per transaction, monthly loan terms at $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$3, $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$4, and $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$5, deposit rate $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$6, and bankruptcy if cash falls below zero after repayments (Zhang et al., 2024). The orchestration framework for financial agents uses execution agents, slippage models $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$7, and audit checks between portfolio and execution stages (Li et al., 1 Dec 2025). In tokenized alternative asset markets, execution is on-chain and compliance-gated: transfers call compliance checks before settlement, tokens can be frozen by governance, and agent stakes can be slashed upon confirmed misconduct (Borjigin et al., 30 Jun 2025).

Some InvestAgent systems remain deliberately non-executable. GuruAgents produces long-only portfolio tables with weights proportional to deterministic scores, rounded to whole percentages, but the paper stops at backtested allocation rather than broker connectivity (Kim et al., 2 Oct 2025). StockBabble is even further upstream: it supports queries, charts, indicator-based recommendations, and simulated portfolio monitoring, but not autonomous trading (Sharma et al., 2021). The concept therefore spans the entire chain from insight generation to full execution and post-trade control.

5. Empirical performance and evidentiary status

Reported performance is generally strong in controlled backtests, though the evaluation regimes differ so substantially that results are not directly commensurable. InvestAlign reports substantial alignment gains relative to pre-SFT models: on $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$8, Overall MSE improves from $$\Delta(P_1,P_2)=\frac{1}{2}\int \big(P_1(t)-P_2(t)\big)^2dt,\qquad \delta(P_1,P_2)=\frac{1}{2}\int \big(P_1'(t)-P_2'(t)\big)^2dt.$$9 to $P3$0 for GPT-3.5, from $P3$1 to $P3$2 for Qwen-2, and from $P3$3 to $P3$4 for Llama-3.1; on the more complex $P3$5 and $P3$6 tasks, similar large reductions are reported (Wang et al., 9 Jul 2025). The agentic REITs system reports mean cumulative returns of $P3$7 for the DeepSeek-R1 version and $P3$8 for the Qwen3-8B fine-tuned version, versus $P3$9 for buy-and-hold; mean Sharpe ratios of $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$0 and $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$1 versus $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$2; and mean maximum drawdowns of $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$3 and $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$4 versus $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$5 (Li, 22 Jan 2026).

In multi-agent trading, headline results are larger still. QuantAgents reports $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$6, $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$7, $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$8, $$\hat{P}_2(t)=\frac{v}{\alpha_2\sigma^2}\exp[r(t-T)],$$9, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$0, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$1, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$2, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$3, and $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$4 on NASDAQ-100 constituents over 2021–2023, and also reports live-trading results of $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$5, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$6 on A-stocks and $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$7, $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$8 on HK-stocks over 24Q3–25Q1 (Li et al., 6 Oct 2025). The orchestration framework for financial agents reports a stock-trading return of $$\hat{P}_1(t)=\frac{\alpha_2\sigma^2\eta e^{2r(T-t)}+\theta_1}{\alpha_1\sigma^2\eta e^{2r(T-t)}+\theta_1}\cdot\frac{v}{\alpha_2\sigma^2}e^{r(t-T)},$$9, Sharpe ratio $\eta>0$0, and maximum drawdown $\eta>0$1 on hourly data from 04/2024 to 12/2024, and a BTC-trading return of $\eta>0$2, Sharpe ratio $\eta>0$3, and maximum drawdown $\eta>0$4 on minute data from 27/07/2025 to 13/08/2025 (Li et al., 1 Dec 2025). FinAgent reports over $\eta>0$5 average profit improvement across six datasets, with a $\eta>0$6 return on one dataset corresponding to an $\eta>0$7 relative improvement (Zhang et al., 2024). GuruAgents reports a $\eta>0$8 CAGR for the Buffett agent on NASDAQ-100 constituents from Q4 2023 to Q2 2025 (Kim et al., 2 Oct 2025).

Enterprise and screening systems evaluate along different axes. AI PB reports Factuality $\eta>0$9, Safety $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$0, Alignment $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$1, inter-rater $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$2, and a hybrid recommender that increases daily feed engagement by $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$3 while reducing repetitive content by $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$4; Shinhan-Guard achieves F1 scores of $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$5 on Toxicity, $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$6 on HarmBench (standard), $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$7 on HarmBench (contextualized), $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$8 on Safe-Guard Prompt Injection, and $$\pi_{\mathrm{FinAgent}}(a_t \mid s_t,\mu_t)=\mathcal{D}^\lambda\!\left(LLM\!\left(\phi_D^\lambda(s_t,\mu_t)\right)\right), \quad \mu_t=\mu(s_t,Mem_t^\lambda,Tool_t^\lambda),$$9 on PII Detection (Park et al., 23 Oct 2025). The agentic screening framework for S&P 500 data reports a best Sharpe ratio of $\mathbf{R}$0 and annualized return of $\mathbf{R}$1 for the full Agentic AI system, versus a benchmark S&P 500 Sharpe ratio of $\mathbf{R}$2 over 2020–2024 (Caner et al., 24 Mar 2026). StockBabble, by contrast, evaluates user-facing rather than financial outcomes: $\mathbf{R}$3 of participants reported feeling more confident investing after use, and all $\mathbf{R}$4 said they would consider recommending the system (Sharma et al., 2021).

Not all evidence is favorable. The empirical study of DeFi investment agents finds that many projects do not provide clear evidence of autonomous trade execution; developer interviews suggest that many visible deployments remain basic API integrations. Across $\mathbf{R}$5 Solana-based agent treasuries covering $\mathbf{R}$6 token holders, treasuries retain about $\mathbf{R}$7M, the top $\mathbf{R}$8 of profitable wallets capture $\mathbf{R}$9 of total gains, and tokens decline $\Gamma$0 on average from all-time highs (Yu et al., 27 May 2026). This is a critical counterpoint to the more optimistic backtests.

6. Limitations, governance, and future directions

The limitations recur with notable regularity. Oracle dependence, model false positives and false negatives, incentive calibration, scalability bottlenecks, and cross-jurisdiction compliance are explicitly identified in the tokenized-asset governance framework (Borjigin et al., 30 Jun 2025). FinArena highlights hallucination risk and information asymmetry, particularly in A-shares, and adds adaptive RAG and human risk-preference conditioning as mitigation rather than final resolution (Xu et al., 4 Mar 2025). AlphaAgents notes that risk-averse prompt conditioning can underperform in strong rallies because conservative exclusions dominate upside capture (Zhao et al., 15 Aug 2025). StockAgent emphasizes leakage avoidance through anonymized assets, sequential prompt injection, and time-consistent information release, implying that many agentic finance evaluations remain vulnerable if such controls are absent (Zhang et al., 2024).

Governance responses have become increasingly formalized. The tokenized-asset architecture embeds continuous monitoring, on-chain incident recording, slashing, and trust re-estimation through an AI Governance Agent loop that computes $\Gamma$1 and can freeze tokens when critical issues are verified (Borjigin et al., 30 Jun 2025). The orchestration framework for financial agents inserts audit agents and memory agents between all critical stages, explicitly barring test-window labels from LLM-facing components (Li et al., 1 Dec 2025). AI PB operationalizes component-level deterministic routing so that any PII-linked workflow is forced onto internal models, with guard-mediated fallbacks and full audit logging (Park et al., 23 Oct 2025). The venture-capital due-diligence framework uses a structural fallback mechanism that writes “Not Found” instead of hallucinated financials, a design choice that directly addresses epistemic uncertainty in high-stakes finance (Alexandrou et al., 13 May 2026).

A broader controversy concerns the meaning of autonomy itself. The DeFi study argues that on-chain observability does not establish decision provenance, because public wallets reveal what happened but not whether trades were initiated by autonomous pipelines or manually signed by humans (Yu et al., 27 May 2026). In response, that paper proposes a maturity framework with three dimensions: autonomous execution, risk-adjusted profitability, and stakeholder alignment. This suggests a likely future trajectory for InvestAgent research: stronger provenance guarantees, richer risk-adjusted evaluation beyond cumulative return, and clearer mapping from system performance to end-user welfare.

Across the literature, InvestAgent is moving toward auditable, multimodal, specialized, and governance-aware designs. Some systems privilege theoretical alignment, some emphasize portfolio optimization, some prioritize retrieval-grounded due diligence, and others pursue direct market execution. The central research problem is no longer merely whether an agent can emit a trade or a stock pick; it is whether the full chain from information ingestion to decision, allocation, execution, and accountability can be made technically robust, empirically credible, and institutionally trustworthy (Wang et al., 9 Jul 2025, Dangi, 2013, Li et al., 6 Oct 2025, Yu et al., 27 May 2026).