"Agents can work superhuman hours, monitor a superhuman number of things, and constantly learn from… mistakes."
That's Michelle Li, co-founder of Spectre Intelligence, describing the pitch behind her startup. The Y Combinator-backed firm claims to be running autonomous AI agents that execute trades across two specialty pods and turning a profit—though that profitability claim remains self-published with no independent validation.
Spectre operates trading desks in semiconductors and biotech, according to its website last accessed in late September. The company describes itself as a "neolab trading firm building an evolutionary agentic system for reasoning under uncertainty." Translation: it's betting that compute power can replace expensive human traders, at least in specialized corners of the market. Spinning up a new pod "costs compute, not headcount," the company noted in its Y Combinator profile from the Summer 2026 batch.
Li, the CEO, is on leave from Harvard, where she studied economics and biology. Her co-founder Joshua Zyzak, the CTO, holds computer science and electrical engineering degrees from the same school. The team is small. LinkedIn lists the company size as 2-10 employees as of earlier this month.
The firm has disclosed neither assets under management nor funding totals, and Spectre's site describes their activities without external performance validation. It does, however, run something called the "Spectre Competition" on a related site, spectretrading.ai. The setup pits a model labeled "Spectre" against flagship large language models in prediction markets, each allocated $1,000. A leaderboard from mid-September positioned Spectre as "beating the S&P 500 by 2.5%," though the leaderboard appears promotional and the methodology remains unverified.
Still, Spectre isn't alone in trying to make this model work.
The New Cohort
Y Combinator's 2026 batches produced a cluster of AI-native trading startups. Standard Signal, from the spring batch, bills itself as a "hedge fund where AI researches and executes every trade end-to-end." AYVID, from the summer cohort, uses similar language. Multiplier, also summer, builds "agent harnesses for asset managers," positioning itself as a toolmaker for institutional desks rather than a fund.
The larger financial world is watching, if cautiously. Numerai, an established quantitative fund, reported around $700 million in assets under management and said it trades more than $1 billion monthly across 30 global markets, according to an August blog post. Man Group, one of the world's largest hedge funds, deployed "AlphaGPT," a proprietary large language model workflow that processes market data at speed. Unlike the autonomous agent model Spectre espouses, Man Group keeps humans in the approval loop.
A Mercer survey fielded in February and published in May found that 55 percent of asset managers had integrated AI into at least one strategy. Another 27 percent were running pilots. Asset managers and private equity firms expect to spend an average of $148 million on AI over the next year, according to a KPMG survey published in April.
So the money is flowing, and the interest is real. The question is whether the technology delivers.
The Performance Gap
Academic research published over the past two years has documented stubborn problems with AI trading agents. Open benchmarks including AI-Trader and StockBench show most agents underperform and exhibit poor risk control, even when they demonstrate strong reasoning on static question-and-answer tasks. A survey synthesizing research on agentic large language models in finance, published in the findings of a major natural language processing conference in late 2025, called for "transparent architectures, stronger evaluation, and risk governance." It cited fragility, weak numerical reasoning, and inconsistent live trading outcomes.
General intelligence, the paper noted, does not equal trading returns. One common failure mode: inadequate risk controls.
Multiple surveys from 2025 and 2026 found that while agentic systems show promise in planning and tool use, audited track records for fully autonomous AI traders in institutional public-equity markets remain sparse. Most published evidence consists of backtests, simulations, or vendor-asserted trials rather than third-party verified performance.
That gap between promise and proof is widening just as regulators start paying closer attention.
Regulatory Scrutiny
U.S. regulators maintain what they call a principles-based approach to AI in trading, which is to say they haven't written new rules yet. The Securities and Exchange Commission withdrew a proposed rule on predictive data analytics in June 2025. As of late September, no prescriptive rule governing AI trading tools is in force. Firms must instead rely on existing fiduciary obligations and anti-fraud regimes.
The Commodity Futures Trading Commission issued a staff advisory in late 2024 reminding market participants that compliance obligations "are technology neutral." Autonomy does not relax regulatory requirements.
In Europe, the EU AI Act entered staged enforcement, with transparency obligations starting in August. Systems deemed high-risk in specific use cases may face future conformity assessments and documentation requirements.
Then came September. OpenAI paused training after agents probed U.S. government sites and accessed non-public files, according to reports from Axios and the Associated Press. Nvidia introduced an Open Agent Safety Platform in response. The Bank of England embedded analysis of the risks presented by AI into their mainstream assessment of risks to systemic firms, as noted in its July Financial Stability Report.
The Financial Stability Board published sound practices in June emphasizing governance, explainability, cyber dependencies, and concentration risks. FINRA highlighted agent governance in its 2026 regulatory oversight report. Both signal that firms with mature AI risk management and resilient agent pipelines are likely to gain regulatory approvals and scale sooner than those treating agents as black boxes.
The Testing Ground
Prediction markets have emerged as something of a proving ground. They saw more than $25 billion in total U.S. trading volume in 2025, according to a Federal Register docket published in June. Spectre's public competition site uses prediction markets as a testing environment, though the firm declined to disclose whether it manages external capital or when it plans to register with regulators if it does.
Li's assertion that "AI is the worst it ever will be" suggests confidence that current capabilities already strengthen the firm's trading principles. Whether those principles translate into sustained, risk-adjusted returns under real market conditions remains an open question. The firm's profitability claim, unverified and unaudited, is the kind of assertion that draws interest in the startup world but skepticism from institutional allocators who have seen plenty of backtests that didn't survive contact with live markets.
For now, Spectre Intelligence is part of a broader experiment: testing whether artificial intelligence can genuinely replace human discretion in one of the most competitive arenas in finance. The results, when they come, will be closely watched.
