One task, one set of permissions
For example: “rebuild the data for this hypothesis as it was known at the time.” Nothing beyond it.
Agentic Investment Management
Agentic investment management, explained step by step.
Agentic investment management is still new to most allocators. This page explains, in plain language, what an agent is, which agents Tensor is designing, what the five research gates do and what always remains a human decision.
Illustrative internal architecture in development · no investment product or regulated service
01 · The basics
What is an agent?
An agent is software that is given one narrowly defined task, access to specific data and tools, and explicit permissions. Within that task it can read, calculate, write and test code and draft a report. It works like a specialist analyst following a checklist, except that every step it takes is recorded.
Prepare, test, record, escalate
It gathers evidence, runs the tests it is allowed to run, writes down what it did and flags anything unusual to a person.
Approve its own work
It cannot change its task, widen its permissions, promote its own conclusion or change risk. Those steps belong to people.
Tensor is designed around several specialised agents rather than one opaque model. No single agent, model or signal controls the portfolio.
02 · The agents
Specialists with separate permissions.
Each agent is designed for one part of the research work. The separation matters: the agent that proposes an idea is not the one that tests it, and none of them can approve it.
- Idea & registration
Research agents
- Designed to
- Scan distinct markets, data families and structural changes, and turn observations into testable hypotheses with fixed rules.
- Cannot
- Test or approve their own idea.
- 02Reconstruct
Data reconstruction agents
- Designed to
- Rebuild point-in-time data: what was known on each historical date, with timestamps, universe and every transformation recorded.
- Cannot
- Use revised data or information from the future.
- 03Challenge
Challenge agents
- Designed to
- Act as an internal red team: search for overlap with known factors, instability and any reason to reject the idea.
- Cannot
- Soften a test after seeing the result.
- 04Validate
Validation agents
- Designed to
- Open a sample kept untouched, once, under the decision rule fixed at registration.
- Cannot
- Reopen the sample or redesign the idea around it.
- 05Implement
Risk & implementation agents
- Designed to
- Model turnover, slippage, liquidity, contract rolls and capacity, and check exposure against volatility, concentration, liquidity and drawdown limits.
- Cannot
- Set or change a risk limit.
- Throughout
Reporting & operations agents
- Designed to
- Prepare recurring reports, route documents, keep the oversight ledger and flag exceptions.
- Cannot
- Communicate externally on their own.
A manager agent routes work between them and escalates to people. The Tensor control plane records permissions, provenance and review status for every step.
03 · The five gates
Follow one hypothesis through the pipeline.
Every idea passes the same five gates in the same order. At each gate it can be rejected, and the reason is recorded. Scroll on, or hover over or focus a station, to see what happens at each step.
Illustrative example with a synthetic hypothesis · not Tensor research data
- IdeaAn observation becomes a written hypothesis.
- RegisterFix the thesis and the rules before any test.
- ReconstructRebuild the data exactly as it was known at the time.
- ChallengeTry hard to prove the idea wrong.
- ValidateOpen untouched data once, under a fixed rule.
- ImplementCheck that it survives real-world costs and limits.
- Human approvalPeople decide. The system stops here.
Current step Idea
- H-0042 · synthetic example
- Most hypotheses are rejected at a gate
Idea
An observation becomes a written hypothesis.
A research agent notices a possible pattern in market data and writes it down as a hypothesis with an ID. At this point it is only an idea to be tested: not a signal and not a position.
- Done by
- Research agent
- Recorded
- Source observation, agent, timestamp
H-0042 · at this stepA research agent observes a possible relationship in the shape of the yield curve and records it as H-0042 · curve momentum.
RecordedRegister
Fix the thesis and the rules before any test.
Before any data is tested, the economic reason, the expected direction, the time horizon and the conditions under which the idea will be rejected are written down and time-stamped. The rules cannot be changed afterwards without creating a new, recorded variant.
- Done by
- Research agent · control plane
- Recorded
- Registered thesis and rejection rules
H-0042 · at this stepThe thesis, expected sign, horizon and rejection criteria of H-0042 are locked.
PassedReconstruct
Rebuild the data exactly as it was known at the time.
Historical data is often revised later, and today’s index members were not always members. This gate rebuilds the data as it was available on each past date, so a test never uses information that was not yet known.
- Done by
- Data reconstruction agent
- Recorded
- Data vintages, universe, timestamps, lineage
H-0042 · at this stepThe curve data for H-0042 is rebuilt from point-in-time vintages; each transformation is logged.
PassedChallenge
Try hard to prove the idea wrong.
Independent agents attack the result: placebo tests, different sub-periods, small changes to the parameters, overlap with well-known factors and alternative definitions. This is where most hypotheses are designed to stop.
- Done by
- Challenge agents (internal red team)
- Recorded
- Every test run, every variant, every rejection reason
H-0042 · at this stepH-0042 is tested across sub-periods, perturbations and its overlap with an existing trend exposure. A hypothesis that merely repeated that exposure would be archived here, with its reason recorded.
PassedValidate
Open untouched data once, under a fixed rule.
A portion of the data is set aside and never looked at during research. It is opened once, and the result is judged by the decision rule registered at gate 01. There is no second attempt on the same sample.
- Done by
- Validation agent
- Recorded
- Access log, decision rule, outcome
H-0042 · at this stepThe untouched sample for H-0042 is opened once; the outcome is judged against the pre-committed rule.
PassedImplement
Check that it survives real-world costs and limits.
Research results can look good on paper and fail in practice. This gate models trading costs, slippage, liquidity, contract rolls and capacity explicitly, and checks the proposed exposure against the risk limits that people have set.
- Done by
- Risk & implementation agents
- Recorded
- Cost, liquidity and capacity assumptions; limit checks
H-0042 · at this stepTurnover, cost and capacity assumptions for H-0042 are modelled; its proposed exposure is checked against the limits.
PassedHuman approval
People decide. The system stops here.
The evidence file goes to responsible people. They decide whether the hypothesis may be used, with which limits, or not at all. No agent can move it further on its own, and rollback remains a human decision.
- Done by
- Responsible people
- Recorded
- Reviewer, decision, limits, date
H-0042 · at this stepH-0042 is waiting for review. The system cannot move it any further.
Awaiting review
After approval: the portfolio layer
- Combine
Approved, independent evidence is combined across strategies and markets.
- Size
Volatility, concentration, liquidity and drawdown limits shape the exposure.
- Monitor
Implementation, exceptions and signal decay stay visible after deployment.
Observed outcomes and decay feed the next research cycle.
04 · What stays with people
The system does the work. People make the decisions.
Agents are designed to prepare, test, reconcile and escalate. Anything that would change risk waits for a person.
Designed for the system
- Scanning markets, data families and structural change
- Turning observations into testable hypotheses with fixed rules
- Rebuilding point-in-time data and attacking every result
- Recording each tested variant and every rejection reason
- Preparing reporting, routing documents and flagging exceptions
Reserved for human approval
- Investment decisions
- Risk limits and exceptions
- Approval, deployment and rollback of models
- External communication
No agent promotes its own conclusion.
05 · What allocators gain
Evidence you can follow, not a black box.
The architecture is designed to answer the questions allocators already ask of a systematic manager, and to keep the answers on record.
Traceability
Each hypothesis is designed to carry an ID, a registered thesis and a recorded path through the gates, including where and why it stopped.
Audit trail
Data vintages, tested variants, assumptions, reviews and decisions are designed to be preserved, not only the winning result.
Less overfitting
Rules fixed before testing, data rebuilt point-in-time, a sample opened once and a count of every variant tried are designed to make it harder to mistake noise for signal.
Control
Agents cannot promote their own conclusions. Investment decisions, risk limits and approvals stay with responsible people.
06 · Glossary & FAQ
Plain definitions.
- Agent
- Software with one defined task, specific tools and explicit permissions. It prepares and tests; it does not decide.
- Hypothesis
- A testable statement about markets, written down before it is tested. Most are rejected.
- Gate
- A checkpoint every hypothesis must pass, in order. Failing a gate stops it, with the reason recorded.
- Point-in-time data
- Data exactly as it was known on a past date, without later revisions or hindsight.
- Overfitting
- A model that fits past noise rather than a durable effect. It tends to look better in a backtest than in real use.
- Untouched sample (holdout)
- Data set aside and not looked at during research, opened once for validation.
- Audit trail
- The complete record of what was tested, by which agent, on which data, with which result, and who approved it.
- Control plane
- The shared layer that records permissions, provenance and review status for every step.
What this is
- An internal research architecture in development
- A way to make research more observable and easier to audit
- A process in which people make the investment decisions
What this is not
- Autonomous trading or a “robo” portfolio
- A fund, product, offer or investment service
- A promise of returns or a performance claim
- A black box in which AI makes the decisions
Frequently asked questions
What is Tensor Asset Management?
Tensor Asset Management S.à r.l. is a Luxembourg company building an agentic liquid alternatives research architecture centred on Tensor-Qube, its internal portfolio research architecture. Specialised agents widen the research search, independent research gates reject what does not survive, and investment decisions, model promotion and risk limits require human approval. The research covers two programmes: multi-asset long/short and systematic CTA. The company is in its setup phase. It is not currently authorised, licensed or directly regulated by a financial supervisory authority, has not commenced any regulated activity and offers no investment product.
Does an AI decide what to buy or sell?
No. The architecture is designed so that investment decisions, risk limits and approvals stay with responsible people. Agents prepare and test the evidence.
Can an agent change the rules after seeing a result?
No. The thesis and the rejection rules are fixed at registration. A change creates a new variant, which is recorded and counted.
Why would most hypotheses be rejected?
Testing many ideas makes it easy to find patterns that are only noise. The gates are designed to reject by default, so that only evidence that survives every check moves forward.
Is this a product I can invest in?
No. This page describes an internal architecture in development. It is not a fund, product, offer or investment service.
Tensor architecture
Agents expand breadth. The architecture preserves discipline.
The objective is not autonomous investing. It is a more observable research and portfolio process in which evidence, constraints and review status remain connected.
Explore the strategiesImportant information
This page describes an internal research architecture in development and is presented solely as corporate and research information. It is not a publicly available fund or investment product and does not constitute an offer, invitation, solicitation, recommendation or performance representation. Tensor Asset Management S.à r.l. is not currently registered or authorised by the CSSF as an alternative investment fund manager and does not currently provide regulated investment services. Any future implementation remains subject to all applicable legal and regulatory requirements and mandate-specific documentation.