Skip to main content

The agentic firm

Agents research and argue. Plain code measures. The founder decides.

Qentari is built as a circuit, not a committee. AI agents do the reading and the arguing, deterministic code does the arithmetic and enforces the limits, and a person keeps the decisions that matter. This page shows each part and says plainly how much of it exists.

BuiltPartly builtPlanned

Checked against code on 17 November 2026 (commit b2af5ec). AI-assisted review of the code (Claude); founder sign-off pending.

Structure

The circuit, stage by stage.

An idea moves left to right. Any stage can stop it; none can wave it through alone. The rule-based models use no machine learning, and no AI system makes an investment decision.

  1. Data

    Built

    Pulls filings, company facts, news, prices and exchange rates into each research run, and records where every item came from.

    What it may and may not do

    It may

    • Read market data from public sources and data vendors
    • Cache results so providers are not called on every run

    It may not

    • Place or change an order
    • Fill a missing source with invented data

    Checked by

    Each data slot records its source and whether it answered; fact and news pulls in a research run are written to the audit log. A failed source stays empty, and the agents downstream are briefed to flag the gap and lower their confidence.

    Human control

    The founder chooses and signs up every data source.

  2. Research agents

    Built

    Fundamentals, technicals, sentiment, macro and quant analysts each write a structured view of the same tickers, in parallel.

    What it may and may not do

    It may

    • Interpret the data they are given
    • Flag a data gap and lower their own confidence

    It may not

    • Compute metrics: their brief is to interpret quant-core results
    • Invent data: their brief is to flag gaps instead
    • Place an order

    Checked by

    Every output must fit a typed schema and carries an evidence level, from raw data to model inference. The Bull, the Bear and the Skeptic read it next.

    Human control

    Every run is in the audit log for the founder to review.

  3. Bull

    Built

    Builds the strongest case for the idea from what the research agents found.

    What it may and may not do

    It may

    • Argue the long case
    • Cite the research agents' findings

    It may not

    • Invent data: its brief limits it to what the research agents produced

    Checked by

    The Bear answers it, and the Skeptic reviews both sides.

    Human control

    Its argument is kept in the audit log for the founder to review.

  4. Bear

    Built

    Attacks the bull case: accounting risk, supply chains, regulation, shrinking margins and macro stress.

    What it may and may not do

    It may

    • Argue the case against
    • Press on low-confidence findings and data gaps

    It may not

    • Invent data: its brief limits it to what the research agents produced

    Checked by

    The Skeptic reviews the full debate.

    Human control

    Its argument is kept in the audit log for the founder to review.

  5. Skeptic

    Built

    A hostile review of every thesis. Its brief lists ten ways an idea can be wrong, from lookahead and survivorship bias to costs, capacity, crowding and "why now", and tells it to reject by default.

    What it may and may not do

    It may

    • Reject, or ask for more data
    • Approve, with its brief requiring a justification for each check

    It may not

    • Approve on prose alone: its brief forbids it

    Checked by

    Its verdict routes the workflow in code. Anything but an approval ends the run: the research manager and execution steps never run.

    Human control

    No agent can overrule a rejection: the run simply ends. The founder writes its brief.

  6. Quant core

    Built

    Deterministic code computes the numbers: risk-adjusted returns, drawdowns, value at risk, position sizing, correlations, backtests and portfolio risk limits.

    What it may and may not do

    It may

    • Compute metrics and backtests
    • Fail a proposed trade against portfolio limits

    It may not

    • Call a language model
    • Change the limits it enforces

    Checked by

    Unit tests pin the maths.

    Human control

    The founder sets the limits it enforces.

  7. Risk veto

    Built

    A deterministic check that every proposed action must pass: kill switch armed, market open, position size, drawdown, daily loss, confidence floor, allocation bounds, approved universe, pre-mortem and walk-forward gates.

    What it may and may not do

    It may

    • Reject an action an agent approved
    • Flag a strategy whose results decay; the founder can switch that one strategy off without stopping the rest

    It may not

    • Approve an action an agent rejected
    • Re-arm the kill switch

    Checked by

    If a rule errors, the action is rejected. One exception: when the approved-universe list cannot be read, the action passes with a warning. A kill-switch drill is built in and is run by the operator.

    Human control

    The founder sets the limits. Re-arming the kill switch is a human action, through the operator controls; no agent code path re-arms it.

  8. Human sign-off

    Partly built

    The research workflow has a built-in pause before its final synthesis, for the founder to review. The shipped entry points (the web interface, and the command line by default) skip that pause. Real capital is not connected.

    What it may and may not do

    It may

    • Review the research record at the pause
    • Trip the kill switch at any time

    It may not

    • Re-arm the kill switch: no agent code path does

    Checked by

    Kill-switch trips and re-arms are written to the audit log.

    Human control

    The founder sets the limits and decides whether real capital is ever connected. Re-arming the kill switch is a human action, through the operator controls.

  9. Execution

    Partly built

    Its brief is to turn an approved decision into an order plan: order type, timing and slicing by size and liquidity. Order slicing is not built as code. Every order passes the risk veto and the kill switch first.

    What it may and may not do

    It may

    • Propose order type, timing and slicing
    • Send orders to live-tracked accounts

    It may not

    • Place an order that breaks a risk limit
    • Act while the kill switch is tripped

    Checked by

    Orders and fills are reconciled with the broker and written to the audit log.

    Human control

    The founder sets the limits it works inside. Real capital is not connected, and only the founder can connect it.

Discipline

Six gates. Most ideas fail one.

A strategy is promoted only after it has passed every gate, in this order. Each gate exists to kill a specific kind of mistake.

  1. Gate 1

    Backtest

    What it kills

    An edge that disappears once decisions use only the prior close, fills come at the next open and trading costs are charged.

  2. Gate 2

    Walk-forward

    What it kills

    Results that do not hold on later windows the idea was never tuned on.

  3. Gate 3

    Skeptic

    What it kills

    No answer to the strongest objection: lookahead, survivorship, costs, capacity, crowding or "why now".

  4. Gate 4

    Risk

    What it kills

    Any breach of a hard limit, or any rule that cannot be checked.

  5. Gate 5

    Simulation

    What it kills

    Fills, turnover or drift on a live-tracked account that stray from what the backtest said would happen.

  6. Gate 6

    Human sign-off

    What it kills

    The founder is not convinced. Nothing is promoted without a yes.

What is real today

Built, partly built, planned.

Every claim about the technology is listed here with its status. If something is not finished, it says so.

Built

Running today, with tests.

07
  • Qentari Models: Alpha Max and Alpha Balanced

    Built

    Two rule-based engines: long-only, trend-filtered momentum in US stocks. They rebalance weekly and check risk exits every trading day. Both run unattended on live-tracked accounts at market prices.

  • Lookahead-free backtesting

    Built

    Every backtest decision uses data through the prior close and fills at the next open, with trading costs charged. Tests fail if that rule breaks.

  • Multi-agent research workflow

    Built

    Specialist agents analyse the same tickers in parallel. A Bull and a Bear argue the case. The Skeptic, briefed to reject by default, reviews the debate. Anything but an approval ends the run.

  • Deterministic quant core

    Built

    Risk and return metrics, position sizing, correlations, backtests and risk-limit checks run as plain, tested code. Agents are instructed to cite its results, not compute them.

  • Hash-chained audit log

    Built

    Every agent call and verdict is written to an append-only log. Each row carries the hash of the row before it, so an edited or deleted row breaks the chain.

  • Kill switch

    Built

    A global switch, plus one per strategy and one per account, halts new orders; a trip also closes positions. It fails closed: if its state cannot be read, it counts as tripped. Re-arming it is a human action.

  • Monitoring dashboard and Qentari World

    Built

    A private dashboard reports each model's health, scheduled runs and orders. Qentari World, a public board, shows where capital appears to be rotating across sectors, regions and asset classes, from free public data.

Partly built

Some of it works; the gap is stated.

05
  • Champion/challenger promotion

    Partly built

    Each model is meant to run only from a tagged release. Alpha Max ran an untagged build from 3 to 12 November 2026; both models run tagged code again. A new strategy version replaces the running one only after walk-forward validation, realistic costs and a shadow run on a live-tracked account, and two challengers are in shadow runs now. The previous tag stays ready for rollback.

  • Human approval gate

    Partly built

    The research workflow has a built-in pause before its final synthesis, for the founder to review. The shipped entry points (the web interface, and the command line by default) skip it, so the gate is not yet part of every run.

  • AI-vs-rules tournament

    Partly built

    The agent desk and a rules-only twin are scored on the same registered window, with every decision on the hash-chained record. The scoring harness is built. The first scored run is still ahead.

  • Point-in-time shadow book

    Partly built

    Alpha Max's exact rules, applied to the current members of two broad US indices instead of a hand-picked list, on a separate live-tracked account. It measures how much of the model depends on its stock list. Three such shadow books now run.

  • Desk agents and conferences

    Partly built

    Equity, crypto, macro, options, futures and risk desk agents run on a schedule and share lessons, budgets and emergency plans through scheduled conferences. They work on live-tracked accounts only.

Planned

Not built. Described so the direction is clear.

02
  • Machine-learning ranker

    Planned

    A learned ranker feeding a deterministic portfolio layer. An earlier attempt failed on data integrity: survivorship and lookahead in its inputs. A rebuilt version, on point-in-time data, is planned and must pass the same gates.

  • Deployment of capital

    Planned

    Qentari's own capital goes only into strategies that have cleared every gate, including a period on a live-tracked account. No account holds real capital today.

Performance

Measured in public, on its own pages.

The rule-based models run in live-tracked accounts with no real capital, and their record is published as it happens. The agent desk has no scored record yet: its first scored run against a rules-only twin is still ahead.

Qentari is in formation. Founded in Cape Town, South Africa; to be incorporated in the United Arab Emirates. Qentari is not a licensed financial services provider. Regulated services would be provided only by or through licensed firms, once those arrangements exist.

No performance figure appears on this page. Live-tracked and backtested results differ from trading real capital and do not predict future results; each figure sits beside its limits on the pages above.

Who decides

One founder, accountable for all of it.

Founder

Juan van Wyk

Next step

Ask how any part works.

The full rule set, the agent briefs and the operational detail are shared after a first conversation and a mutual NDA.

Live-tracked: broker accounts trading at live market prices with no real capital. Actual results would differ because of fills, slippage and capacity, and this is not a record of managed client money. Past performance does not guarantee future results.

Backtest: simulated, net of 6 bps per side costs, not results of actual trading.

Nothing on this site is an offer or a recommendation. Risk disclosures