Investment agents in 2026: adoption is wide, proven results are few

Adoption has outrun proven results

In under three years the question has changed. In 2024 it was one of permission: 86% of hedge fund managers in AIMA’s survey gave their staff some access to generative AI [12]. In 2026 it is how many agents are running, with what permissions, and who reviews their output. Balyasny’s platform “supports thousands of autonomous agents working 24/7” [1], and JPMorgan’s chief analytics officer, Derek Waldron, says: “We’ve entered now the era of long-running autonomous agents” [2].

Access has become a commodity. Each of the three big labs sells a ready-made financial package [13][14][15], and data providers pass their licensed data to these models through the Model Context Protocol (MCP), the customer’s entitlements travelling with it [16][17]. Model, data and tools are a layer any institution can buy.

What is scarce is proof that an agent does investment work well. The best model on Vals AI’s Finance Agent benchmark scores 65.40% on a junior analyst’s questions, and 55.79% when every check must pass [4]. The Financial Stability Board (FSB) finds “little evidence at present” of generative AI used for fully autonomous trading [10].

The value observed is operational: 69% of Mercer’s managers report better efficiency [5]; among data buyers in Neudata’s survey the tangible return came from efficiency for 41% and from investment performance for 17% [19]; and only 3% of 847 use cases in ESMA’s survey serve investment services [20]. Adoption has run ahead of value, and the bottleneck is measurement.

The report defends three claims:

• Agents have spread upstream of the decision, and the decision has stayed human: the agent reads, analyses and prepares, and a person or a committee accepts.
• The measured gain comes from the institution’s own evidence: evaluation on its tasks, data labelled by its experts, and written controls.
• Execution authority is the open frontier: brokers grant it to individuals in ring-fenced accounts, institutions stop at proposing, and regulators are writing the conditions for crossing it.

What an investment agent is

An investment agent, here, is a system built on a language model that plans multi-step work and carries it out with tools and data, within set permissions and without real-time supervision, then hands over a research package, a recommendation or an order. It is not a systematic quantitative model applying rules or a trained statistical model, like Bridgewater’s [21]; nor a robo-adviser allocating a portfolio by fixed rules; nor an assistant answering a question.

Method, and grades of evidence

The report rests on institutions’ and providers’ announcements, journalism, published research and benchmarks, and regulatory documents up to 11 October 2026, and describes trading results without return figures. Each source carries a grade of evidence in the list of sources, and the grade belongs to the claim, not the publisher: a bank’s annual report is a filing, but what it says about its own tools is the bank speaking about itself. Autonomous means working without real-time supervision; third-party, measured by someone other than the owner.

Citations of grades D and E outnumber official records more than twice over: most of the industry’s impact figures are announced by those with an interest in them, and few are records anyone can check.

No large institution has handed a language agent the portfolio decision

The grid puts each institution where its own announcements place it: six levels of autonomy against six functions of the investment chain. It maps onto the ladder of human oversight in IOSCO’s supervisory toolkit [11]: up to L2 is “human-in-control”, L3 “human-in-the-loop”, L4 “human-on-the-loop” and L5 “human-out-of-the-loop”.

The grid resolves the best-known exception, Bridgewater: its machine-learning system is autonomous in the portfolio decision alone, while risk and execution stay in human hands [21]. Above the line sit rules-heavy work in large institutions, such as routing email at Barclays [29], and retail trading in ring-fenced accounts at brokers [36][38]. Full autonomy in the decision is announced by a quantitative fund running on machine learning, or by small funds not yet confirmed in operation [25][41][42].

Architecture and specialisation move the result

The result moves with architecture and specialisation more than with the choice of model: in a live experiment the agent’s framework shaped its behaviour more than its model [8], and fine-tuning on experts’ judgements produced the clearest published gain [9]. Whoever builds an agent faces five decisions:

Two open cases: Trading-R1 and RD-Agent(Q)

Trading-R1, from UCLA, Stanford and the University of Washington with Tauric Research, trains Qwen3-4B in three graded stages: the structure of an analysis, then claims and their evidence, then the decision. Each stage is supervised fine-tuning followed by reinforcement learning (GRPO), on 100,000 samples from 18 months and 14 tickers [49]. It was backtested over three months of 2024 on five stocks and an index fund, against other language models rather than the market, in a window inside its data period; its weights are unreleased, and its authors recommend it as support for an analyst, not as a decision tool [49].

RD-Agent(Q), from Microsoft Research Asia, appeared at NeurIPS 2025 with its code. It automates the quantitative research cycle: hypotheses, then code an agent writes and tests, then feedback, with a scheduler choosing each round between improving the factors and improving the model. In backtests it beat classical factor libraries with about 70% fewer factors; its main test period, 2017–2020, lies inside what its models already know, so the authors added a test on 2024 and 2025 [52]. It is the open counterpart to Man Group’s AlphaGPT.

Fine-tuning on experts’ judgements

The clearest measured gap in the industry comes from a single case reported by its two parties. On 30 June 2026 researchers from Bridgewater’s AIA Labs and Thinking Machines Lab compared a fine-tuned model with frontier models on six real information-filtering tasks, such as judging which articles and central-bank documents deserve an investor’s attention [9].

The fine-tuned model made 29.8% fewer mistakes than the best frontier model, at a fraction of its cost. The lesson is in the route: the first labels came from non-expert labellers, were often wrong and produced weak models, so the disagreements went to Bridgewater’s experts; automated prompt optimisation added nothing [9]. The limits are plain: one case, by two companies with products to sell, “isn’t a truly independent comparison”, in The Decoder’s words [53], on filtering and labelling tasks, untested on a portfolio decision.

What ownership means

• Owned weights: an open model the institution fine-tunes on its experts’ judgements, keeping its weights and data, as Bridgewater did [53]; Thinking Machines released its open model Inkling for this use [54].
• Licensed weights on the institution’s own infrastructure: HSBC runs Mistral AI’s models on its own systems [55].
• An owned orchestration and evaluation layer over rented models: Balyasny owns BAMAgent and its evaluation suite, and swaps the models beneath them [56][1].

The labs’ financial packages offer the third meaning only: they promise not to train on the client’s data, and by 11 October 2026 none offered fine-tuning on the client’s experts or weights the client owns [13][15][14].

Cost is an engineering variable

On the Vals benchmark the leader costs $4.38 per test, and Muse Spark 1.2 comes within five points of it at a sixth of the cost [4]. In its first version, returns to spending fell away beyond about a dollar a question [57], and at Bridgewater the fine-tuned model cost 13.8 times less per task [9]. Cost per task belongs in every comparison.

Agents on-chain: transparency by design

Hyperliquid is “the densest agent surface in the perpetuals market”, in the words of Alchemy, which sells agent infrastructure [58]. Each vault’s page shows any visitor its profit and loss, open positions and trade history [59], and newer vaults delegate trading to authorised agents with their accounts on-chain [60]. Transparency by design, not evidence of skill.

Live records are few, and the longest lags its index

The strongest evidence about a system that manages money is a published live record: a daily price, a fund’s report to its shareholders, a filing. Such records are few, most belong to machine-learning models rather than language agents, and the table tells them in words.

The closest to a language agent is LIVR. Its prospectus calls the three models the fund’s AI investment committee and leaves the final say to a human analyst [39]; the models, per Bloomberg, are ChatGPT, Gemini and Claude [67]. The fund is small, about $24 million [40], and its third year, in September 2027, is the first reasonable point to judge it.

Survivorship bias is visible: QRFT was liquidated in July 2026, and a fair table names the closed beside the open. Bridgewater says its fund raised $1.6 billion at the start, less than the $2 billion reported [68]. Since March 2026, robo-advisers in Saudi Arabia must publish their portfolios’ performance since inception after actual expenses [69]; rules like it will multiply such records. A record is read net of fees, since inception, against its benchmark, and one shorter than three years is too short to judge.

General scores do not predict trading results

Two results recur in the live experiments published so far: a model’s general score rarely predicts its trading result, and the agent’s framework shapes its behaviour more than its model. Across Alpha Arena’s seasons, the winner changed.

The founder of Nof1, which runs Alpha Arena, says language models “can’t really make money by themselves” [74]. The backtest claims of multi-agent frameworks are far larger, and re-testing one of them over a long period reversed the sign of its return [44]. These experiments are too short and too rarely repeated to separate skill from luck, and no third party audits them: what is needed is a test of the agent’s behaviour, not of its accuracy alone.

Memory manufactures false performance

Does the model reason, or remember?

The strongest objection to backtesting language models is simple: the model has read the history. GPT-4o recalls the S&P 500 on particular dates, and unemployment rates to a tenth of a point, before its knowledge cutoff, and masking names and dates does not stop it [76]. “Lookahead Propensity” is clearly positive before the cutoff and “collapses essentially to zero” after it [77], and most financial agents’ strong backtest returns vanish once the model’s knowledge window ends [48].

The best-known evidence on the other side was withdrawn. A 2024 paper from Chicago Booth found GPT-4 better than analysts at predicting the direction of earnings from anonymised statements, without relying on memory. It was withdrawn in February 2025, after one of its authors found inconsistencies in its data while trying to replicate it [78].

The nearest remedy is a chronologically consistent model: ChronoBERT and ChronoGPT are trained only on text available up to each date, and their authors found look-ahead bias modest in the one application they tested, and specific to model and application [50].

Long tests settle the direction. Over two decades and more than 100 stocks, the advantages reported for language-model strategies “deteriorate significantly” [79]. The March 2026 survey names five evaluation failures: look-ahead bias, survivorship bias, backtesting overfitting, transaction cost neglect and regime-shift blindness [44]. A September 2026 review finds that the public evidence shows no general AI architecture delivering persistent net alpha across regimes [80]. A backtest inside a model’s training window is not evidence of skill.

Man plus machine

The strongest peer-reviewed evidence for combining people and machines comes from traditional machine learning, not a language model. Cao, Jiang, Wang and Yang’s AI analyst beat analysts in 54.5% of stock-return predictions between 2001 and 2018, while people won where institutional knowledge mattered, such as intangible assets and financial distress [81][82]. The combined model avoided about 90% of analysts’ extreme errors and about 40% of the machine’s [82].

Maps of the literature

• Ding and colleagues (March 2026): trading agents are traders or alpha miners, and traders are driven by news, reflection, debate or reinforcement learning [83].
• Dong and colleagues (EMNLP Findings 2025): five financial domains, more than 30 benchmarks [84].
• Chen and colleagues: evaluate a financial agent first on its risk profile, not on point accuracy [85].

Benchmarks stumble on compound work

Vals AI’s Finance Agent benchmark, updated on 7 October 2026, holds 927 expert-reviewed questions on listed companies’ filings in nine categories, and grades with partial credit and by whether every check passes [4].

Gemini 4 Argon leads with 65.40%, and with 55.79% when every check must pass [4]. The gap between categories is wide: 84.81% in earnings analysis, 34.52% in financial modelling. The page’s own text was not updated when the leader was added, and still says no model reaches 51% under strict grading; the first edition of this report repeated that figure [4]. In Vals’ words, models “still struggle to perform reliably on harder, multi-step financial work” [4].

Mercor’s APEX-Agents, which Mercor sells, measures an investment banking analyst’s work in 160 cases; the top score is 71.3% at first attempt, and its list of failures is more useful [86]:

• A rounded figure fails its criterion even when the reasoning is right.
• Errors compound: in the sample task, the sensitivity analysis passes only if all eight of its outputs are right.
• Agents edit and extend existing models, but “reliably building a complex model from scratch remains beyond current agents”.

In retrieval, web search and financial plugins “substantially improve” results [87], and models are more accurate on larger companies yet more prone to invent their figures [88].

Arabic financial reasoning

SAHM, at ACL 2026, holds 14,380 expert-verified instances in seven tasks, from AAOIFI’s Shari’ah standards and fatwas to accounting and event–cause reasoning [89]. Its 19 models are “substantially stronger on recognition-style tasks than on generation and causal reasoning”, and “Arabic fluency does not reliably translate to evidence-grounded financial reasoning” [89]. It tests no agent using tools.

Forecasting: real progress under clean measurement

Where measurement is live and clean, progress is real. In July 2026 Cassi AI became the first system to rank above the superforecaster median on market questions, by a margin below statistical significance [90]. In Metaculus’s competition the bots’ gap to the Pros narrowed from −20.03 points in the second quarter of 2025 to −1.25 in spring 2026, without significance [91].

Under the strictest grading the best agent gets just over half of a junior analyst’s questions fully right, and its weakest points are the core of investment work: modelling, reconciliation across documents and multi-step work.

Evaluation on the institution’s own work makes the difference

Bridgewater’s pattern appears elsewhere in other forms. Balyasny tests every new model on thousands of tasks with verifiable outcomes inside its own environment, and picks the model for each task by its result [1][92]; on the relevant subset, the new model scored 89.4% against 86.1% for the previous one, by its own account [1].

In CPP Investments’ experiment on reconciling a $522 million portfolio seeded with hidden errors, direct prompting left the models disagreeing and missing errors; requiring a confidence score for every finding, with human review below 70%, raised accuracy by 45%, according to the fund [35]. The accuracy came from governance.

The Financial Stability Board writes the same principle: material uses “may justify higher performance requirements and a lower tolerance for errors” [10]. What matters to an institution is how this agent performs on this task with this data, and the answer lies in its own evaluation, not in a public leaderboard [7][4].

The evidence ranks the methods: experts’ data on the institution’s own tasks first [9], then tools and retrieval with verification [35][87], then disciplined backtesting, which exposes false gains more than it creates new ones [79]. This chapter’s own evidence is among the report’s weakest by grade, institutions and vendors speaking about themselves, and its thesis awaits a third-party test.

Institutions automate what comes before the decision

Every kind of institution follows the same pattern: an autonomous agent within the task, and a person at the decision.

Hedge funds and quants

Quantitative and multi-strategy hedge funds are the furthest ahead and the most forthcoming. At Balyasny a merger-arbitrage deal package that took 3–5 days now takes less than one, under written governance: least privilege, permissions per tool, human review of every material output, and “Investment judgment and accountability remain with people” [1]. Man Group names AlphaGPT’s failures, among them “drift” from the idea conceived, and says: “we can’t leave it unsupervised just yet” [23].

At Citadel “nearly all” equities investors use an assistant, and investment judgement stays in human hands [3]; its founder, Ken Griffin, said in October 2025 that generative AI “just falls short” of generating excess returns [98].

High-Flyer, the quantitative fund that built DeepSeek, runs AI in its multi-billion-dollar portfolio [25]. In March 2025 Reuters reported that Baiont Quant, Wizard Quant and Mingshi were stepping up their AI research, and that more than twenty Chinese fund companies, China Merchants Fund and E Fund among them, had deployed DeepSeek locally [25].

Asset managers and banks

Asset managers follow two stages, as Schroders puts it: assistants on public data, then agents on the firm’s own data acting “first with human approval for each action” [99]. BlackRock runs its research platform Asimov without publishing its effect [27], and J.P. Morgan Asset Management’s internal platform now advises the portfolio managers who vote, in place of outside proxy advisers [100]. T. Rowe Price tells two cautionary stories: a model that saw nothing unusual in First Brands’ prospectus while a portfolio manager suspected off-balance-sheet financing, and “all the LLMs” recommending buying the dip as a war broke out [101].

Banks have put a language-model workspace before most of their staff, by their own figures: about 250,000 people at JPMorgan, half of them daily users [26], more than 180,000 at Citi [102], about 100,000 at UBS [103]. Agent platforms are being built on top, such as Citi’s “Arc”, with “a human always in the loop” [102], and agents for trade accounting and client vetting at Goldman Sachs [30]. The model provider is replaceable; the governed data layer is the asset.

Sovereign and pension funds

Sovereign and pension funds redesign operations and keep judgement human. NBIM screens every company on its first day in the portfolio and “always” reviews the information before a decision [34]; it puts its productivity gain at about 20%, “equivalent to 213,000 hours”, an estimate from staff surveys [104]. GIC lists “a virtual Investment Committee member” that challenges proposals, without publishing its design or effect [33].

The Public Investment Fund says AI is “embedded across every layer” of its work: private-market valuation, daily investment reporting by a language model, analysis of asset managers [32]; in 2025 it deployed 15 agents serving more than 2,300 users internally [31]. The Abu Dhabi Investment Authority’s quantitative research team exceeds 125 professionals [105]. In everything announced, asset allocation stays a human decision.

Execution authority is the open frontier

Retail brokers were the first to give agents execution authority publicly. Since May 2026 Robinhood’s agents trade only the balance in a separate account [96]; by September more than 150,000 of nearly 29 million customers had opened one [36][106], and customers bear all the risk [107]. Public’s agents run a plan the user approves, within the customer’s conditions [37], and eToro’s trade a sub-portfolio through a scoped key [38]. Charles Schwab plans a paid agentic option “with strict guardrails” [108], and Vanguard is planning personalised advice in its Digital Advisor for early 2027 [109].

In institutions, authority stops at preparation: firms are “exploring” agents that prepare execution instructions [10], Anthropic’s templates wait for the user’s approval before the work “is acted on” [110], and Google offers hedging “suggestions” [14]. The furthest design published is research: in “The Self Driving Portfolio”, 44 agents build portfolios with 21 methods, a higher agent rewrites their code, and an investment policy statement governs the chain [111].

The paradox: with individuals the agent executes and its risk falls on the customer alone [107]; in institutions accountability surrounds an agent that does not execute. The missing link is a verifiable record of what the agent decided and why, written before the outcome is known. The nearest thing in the regulatory documents is documenting the reasoning at each decision point [10].

The infrastructure is now shared by everyone

The three labs sell nearly the same bundle: a model, MCP connectors to data providers, and agent templates with per-tool permissions and an audit log [13][110]. OpenAI trained its models on junior bankers’ work with more than 100 former bankers paid $150 an hour [112], and its tool cites down to the table and the passage [15]; Google’s financial agent shows its reasoning with confidence scores [14], and Copilot in Excel shows what it will change before it acts [113].

Most large data providers chose to be the licensed layer under everyone’s agents: LSEG with the three labs and Microsoft [114], FactSet with a production MCP server [17], Moody’s with more than 50 ready-made agents [115]. Bloomberg is the exception, keeping its agents inside the Terminal [116]. Spending on market data hit a record $49.2 billion in 2025 [117].

Venture capital goes to automating research: AlphaSense at a $7.5 billion valuation [118], Rogo at $2 billion [119], Model ML in talks above $1 billion [120]. What touches the portfolio decision is small and experimental, such as Boosted.ai’s “multi-model AI investment committee” for select partners [121].

The infrastructure has settled into three standard layers: MCP for licensed data, “skills” for packaging a method, and managed runtimes for permissions and audit [110][14][113]. A newcomer can buy all three, and gains no edge from them.

Calibration, sameness and robustness are still open

Calibration first. On 300 questions resolved after the models’ cutoffs, KalshiBench found all five models “systematically overconfident” [122], and Prophet Arena finds markets absorbing information faster than models as resolution nears, and that is the situation closest to trading [123]. Calibration is measured on the institution’s own questions, not assumed.

Then sameness. Frontier models behave in correlated ways, more so as they grow more capable, creating “a non-diversifiable risk floor”, in a small simulated sample [124]. In a simulated market, reinforcement-learning traders sustained supra-competitive profits without agreement, at the cost of liquidity and price informativeness [125]; T. Rowe Price saw every model it tested recommend the same move [101]. IOSCO names “herding behavior/correlation/collusion” [11], and Sarah Breeden, a deputy governor of the Bank of England, warns that similar agents “could amplify volatility in stress” [126].

Then robustness: unauthorised actions that are hard to undo, errors in any of “hundreds of intermediate steps”, “memory poisoning” [10], and “misaligned incentives” from poorly designed reward functions [127]. An agent that reads documents from outside the institution needs adversarial testing before its permissions widen [1].

Regulators ask for limits, a record and a kill switch

As of 11 October 2026, none of the major regulators we reviewed has a binding rule for investment agents; their expectations sit in guidance, inspection tools and consultations, converging fast:

Enforcement so far punishes overclaiming and neglect: a 2024 settlement with two advisers over false claims about AI [134], and $90 million paid in 2025 by two Two Sigma advisers over known flaws left for years in their models [93]. The second is the closer precedent for agents: a known flaw left unrepaired breaches fiduciary duty, whatever the technology is called.

The regulatory question is moving from “did a person approve every decision?” to “are the limits enforced and the oversight effective?”, and the larger risk is “governance drift”: autonomy growing without oversight to match [135].

The thesis can be tested by 2028

Each indicator tests a part of the thesis:

Scenarios to 2028

• Continuation: agents prepare and propose in institutions and execute for individuals only; its sign, no progress in the benchmarks’ compound categories.
• Institutional crossing: execution authority under written limits, with published evidence cards; its sign, a language agent’s live record ahead of its benchmark after fees over three years.
• Reversal: a market incident tied to similar agents returns a person to every decision; its sign, rising correlation between models’ decisions in live experiments.

Evidence comes before authority

If the evidence is read rightly, these are its lessons for whoever builds, fine-tunes or buys an investment agent:

• Measure on the institution’s task: a public leaderboard rarely predicts performance on it [7][4].
• Let its experts set the right answer: in the clearest published experiment, non-expert labels made weak models [9].
• Freeze the data at the decision date: what a model memorised before its cutoff makes performance that fades after it [76][48].
• Write the limits before the work: regulators ask whether limits are enforced, not whether a person approved every action [126][10].
• Seal the judgement before its outcome: a record written before the outcome is the missing link.
• Start with one task: every measured gain came on a narrow, named task, and they do not add up [9][1][35].
• Differ from the others: frontier models behave alike [124], and fine-tuning on the institution’s experts may soften that, without curing it.

A protocol for live testing

A live experiment, by an institution or a provider, can be read against eight conditions:

• Data after each model’s cutoff, or a chronologically consistent model, with the date stated [50].
• A model version pinned by its identifier, with instructions and architecture frozen throughout.
• Each decision’s fingerprint published, or deposited with a third party, before the outcome is known.
• Simple baselines: buy and hold, the index, equal weights [75].
• Transaction costs and slippage, stated [44].
• A period crossing at least two market regimes, or a statement that it does not [79].
• Several runs with confidence intervals [4].
• Full logs published: inputs, tools, reasoning, decision.

Where we stand

Quantaup is a technology company in its research and development phase. We build a system of investment agents and use it internally today, and we are opening a research partnership to a limited number of licensed institutions: the institution brings a task from its own work, its documents, and what its experts consider the right answer; its data and every decision remain its own.

The industry bought access in 2025 and 2026. The next stage belongs to whoever holds the evidence on the institution’s own work.