Autonomous Agentic Hedge Fund

enter spyglass app

access to the tengen spyglass app is through invitation only

01 · the agents

A fund is not one model. It is a committee.

A hedge fund is a team of competing perspectives with different inputs, different priors and different goals, presenting competing views to a committee that converges on risk-managed positions. Tengen replicates that structure in software: every seat at the table is an autonomous agent, and the disagreement between them is the edge.

the perspectives

macro strategists, fundamental analysts, tape readers, quants

Many autonomous agents, each with its own opinion, thesis and risk appetite, each deliberately blind to what the others see. An agent is a strategy with a time policy, a model that scores what will move, and an algo that decides what to do with money — named, versioned, and carrying its own track record.

the committee table

views argued side by side

The agents trade and argue over the same evidence. Disagreement is preserved long enough to be examined; an eight-for, two-against vote is information, and so is unanimity. Nothing is averaged away and no opinion is discarded — the shape of the vote is the signal.

the minutes

every decision accountable

Every position traces to the views, the versions and the beliefs that drove it, scored against what happened. Every fact carries when it was true and when it could have been known, so every replay shows exactly what an agent could have seen; every number traces to the run, the config and the code that produced it.

02 · the arena

Every agent is judged the same way.

An agent’s opinion earns nothing by itself. Every agent passes through the ring of fire — three standings, one at a time, forward only, on evidence — and enters a ring only on the evidence of the one before. A book that looks good in a backtest is usually garbage, and the arena’s job is to say so, to every agent alike.

  1. 01

    backtest

    the promise

    Event-driven, session by session, over the products as they stood at each decision time. Orders fill at the next open with slippage from the bar and a participation cap; a fill that cannot happen does not happen. Every model is refit walk-forward with an embargo the size of its horizon.

  2. 02

    paper trading

    the record

    A real clock, real data as it arrives, and a balance sheet the engine keeps — cash, positions, orders, fills, marks — stepped one session at a time with no knowledge of the future. Every order and fill is a row. The paper book is the agent’s real record.

  3. 03

    live test

    promise against record

    The agent’s backtest re-run over its paper book’s own dates and set beside it; the divergence decomposed into late data, slippage, unfilled orders, signal drift and a residual, with an alarm when it leaves the band the backtest’s own variance predicts. Nothing places a real order.

the multiple-testing tax

Every backtest run of an agent counts as a trial, and every sweep point is one. The reported Sharpe is deflated for all of them: a good number found by searching is taxed by the search.

the deflated Sharpe

The probability that the true Sharpe exceeds zero given how many trials were run, with the return series’ skew and kurtosis in the denominator. A probability, so its floor is one number under one.

a sweep is a surface

A grid of policy variations, one backtest per point, shown as a surface with the declared point marked and the best point named — never a best point alone.

stability and capacity

Performance by year, by regime, by half of the sample, by a bootstrap of the trades; an edge that sits in one year is labelled so. The same algo at ten and a hundred times the capital, so the number carries its scale.

03 · the window

Spyglass. Watch the committee work.

An invited reader watches the fund through spyglass: what the committee has published, one release at a time, with a note and a timestamp. Start at trading, pick an agent, and follow its book back along the chain to the evidence it argued from.

  1. cataloguewhat evidence exists
  2. pipelineshow it is gathered
  3. curationhow it becomes products
  4. analysesthe questions asked of it
  5. modelsan agent’s opinion
  6. algoswhat the agent does with money
  7. tradingthe agents’ books, read as money

shows

  • every agent’s paper book beside the others, each judged on its own, never added into a blend
  • promise against record: the backtest’s path beside the book’s path, and the divergence decomposed
  • the lineage of every number: a book back through its algo, model, strategy, analyses, products, pipelines
  • frozen reports, per release: what was reported on a date is what the row says, kept unchanged

never

  • raw or curated vendor rows, bars, the wide table
  • orders, fills, the marks themselves
  • a change to anything the committee owns — the window is read-only

Many opinions. One arena.

The agents disagree by design; the platform judges every one of them the same honest way and publishes the result. Nothing real trades before a written deployment gate is passed. What is published is each agent’s book, its promise and its record, and the lineage of every number.

enter spyglass app