The cleanest expression of where agentic trading is going lives in a Sydney hedge fund called Minotaur Capital, which operates with no human analysts at all1. Its proprietary system, Taurient, reads roughly five thousand news articles a day and outputs trade theses; the human team are supervisors, not researchers. Minotaur is a public outlier. The private reality across the multi-strategy and global-macro segments is that something architecturally similar is being built into established firms — not as a single omnibus AI, but as a topology of cooperating agents, each replicating a function that used to live in a named human role.
This note describes that topology. It then identifies the three places where the topology, as it is being built in 2026, is most exposed to the next supervisory cycle.
The analysis trinity
The dominant pattern is what one industry summary calls the analysis trinity2: three specialist analyst agents working in parallel, each on a distinct epistemic basis.
- Sentiment agents. Read news, social media, regulatory filings, transcripts, and adjacent unstructured signals. Output is a directional read — bullish / bearish / neutral — with a confidence interval and a citation chain back to the underlying texts.
- Fundamentals agents. Parse company filings, peer-comparable financials, supply-chain disclosures, and management commentary. Output is a written valuation thesis with explicit assumptions.
- Quantitative agents. Apply statistical and machine-learning models to time-series data, factor exposures, and cross-asset relationships. Output is a numerical forecast with confidence bands.
In the architectures I have reviewed, these three agents do not vote. They produce structured outputs against a common schema and pass them upward to a synthesis layer — sometimes a fourth agent, sometimes a human portfolio manager — that assembles the trade. The trinity exists not because three opinions are better than one, but because the three epistemic modes are mutually un-substitutable: a sentiment read is not a substitute for a fundamentals read, and a quant signal is not a substitute for either.
The risk layer is the hardest part
The supervisory weight of these architectures sits in the Risk agent layer, not the analyst layer. A Risk agent in production typically has:
- Hard authority. It can veto a trade pre-execution and unwind a position post-execution; the analyst agents cannot override it.
- Independent inputs. Position-sizing logic based on volatility regimes, liquidity scoring, drawdown caps, factor-exposure limits, counterparty concentration, and venue concentration.
- A separate decision log. Risk-agent decisions are logged at a different granularity from analyst-agent decisions and are reviewed on a different cadence.
- An override path. Where a human risk officer can pause the agent, narrow its tool surface, or dial down its capital authority without taking the entire system offline.
In firms that have built this layer well, the Risk agent is the most heavily evaluated component of the system — eval coverage is higher, red-team frequency is greater, and the documentation is the most mature. In firms that have not, the Risk agent is treated as a stop-loss script with delusions of grandeur, and the analyst agents have substantially more authority than the firm’s own internal protocol records.
A useful diagnostic question, in due-diligence work: show me the last five Risk-agent vetoes, in writing, with the analyst-agent reasoning that triggered each one. Firms that can produce that document quickly are firms whose Risk layer is real. Firms that cannot are firms where the risk layer is a label.
The execution surface
Below the analyst trinity and the Risk layer sits the execution surface. The architectural decisions here are quieter but equally consequential:
- Order-routing authority. Whether an execution agent selects venues from a pre-approved list or is permitted to discover venues at runtime.
- Order-type authority. Whether the agent issues market, limit, stop, or more exotic order types — and whether a regulator’s definition of automated quoting or high-frequency trading is triggered by the agent’s behaviour even when no human intent existed.
- Venue-side surveillance. Whether the firm’s registered status under CSA, SEC, FCA, or MAS rules implicates additional obligations when the agent is the one originating orders.
These are the surfaces where agentic stops being a software-architecture word and becomes a regulatory question. The CSA Staff Notice on AI use, CIRO Guidance Note 3300, and the FCA’s 2024–2025 AI strategy materials all give examples that map cleanly onto the execution-surface configuration. The harder question is not whether obligations apply — it is whether the firm has documented its agent’s execution-surface configuration in language a registered compliance function and an external supervisor can both read.
Three exposures
Three patterns recur in the architectures I have reviewed in 2026.
First, the harness-monoculture exposure. Many firms are building the analyst trinity on top of the same one or two open-source agentic frameworks, with light modification. This is rational: those frameworks are well-tested and the alternative is reinventing orchestration. But it means that under stress, the trinity at firm A and the trinity at firm B may share enough harness-level architecture to behave correlatedly — even when the firms believe their alpha is differentiated. This is the architectural mechanism behind what the Bank of England is naming as herding3.
Second, the audit-trail latency. Decision logs are typically structured for retrospective debugging, not for supervisory review. The artifact a regulator will eventually request is a per-trade narrative — what the firm decided, why, on what input, against which constraint — readable in plain language by a non-technical reader. Most firms can produce this with a week of work; almost no firm can produce it in the four hours that a stress-event request will allow.
Third, the model-update surprise. The agents in production were tested against a particular underlying model snapshot. When the underlying provider issues an update — improved capability, refined safety behaviour, changed reasoning style — the firm’s evals frequently are not re-run before the update reaches production. The model behaves slightly differently; the agent topology behaves slightly differently; and the firm discovers this in production rather than in evals. Firms that have a model-pinning policy and a model-update review process are rare; firms that need one are all of them.
The supervisory direction of travel
Each of the three exposures named above has, in the public 2026 record, a supervisory authority asking a question about it4. The Bank of England is asking about herding. The CSA is asking about audit trails and accountability under model-mediated trading. The SEC and CFTC are asking about authorization granularity and the definition of automated trading. The FCA and MAS are tracking adjacent questions on operational resilience.
The supervisory direction of travel is consistent: written documentation of the agent topology, the authorization surface, and the decision log will become the operative supervisory artifact. Firms that build that documentation as a byproduct of their architecture are well-positioned. Firms that treat documentation as a downstream compliance task will rebuild parts of the system to make documentation possible.
The Risk agent, the execution surface, and the harness choice are not just technical decisions. In 2026 they are documentary commitments to a supervisory frame that is now being written.
Notes and citations
On Minotaur Capital and similar AI-native funds: see public reporting in 2025–2026 on AI-only investment management firms. Minotaur is referenced here as a public exemplar of the architectural direction; specific performance figures are reported by the firm itself and have not been independently verified for the purpose of this note. ↑
The “analysis trinity” framing — Sentiment, Fundamentals, Quantitative — appears across multiple practitioner write-ups of agentic hedge-fund architectures published in 2025–2026. ↑
Bank of England Financial Policy Committee record, April 2026; see the companion working note The Bank of England names the agentic-herding problem. ↑
Survey of public regulator statements 2024–2026: CSA Staff Notice 11-348; CIRO Guidance Note GN-3300; SEC and CFTC public remarks on autonomous-system supervision; FCA AI strategy 2024–2025; MAS technology risk management guidelines update.
AIMA member surveys 2024–2025 record substantial year-on-year growth in employee access to generative-AI tools across surveyed hedge funds.
On model-pinning and update review: practitioner observation; few firms have written policies in 2026.
On harness-level monoculture as a stress-correlation mechanism: a research direction worth its own working paper.