Case study · Self-learning

Intelligence that compounds — in weeks, not years.

A human analyst takes years to build market judgement. AG-3 — three specialist agents and a synthesising boss — accumulated it trade by trade across a ~40-session backtest: underperforming at first, catching up, then beating the market in the final stretch.

The learning curve, compressed

Behind at first. Level by mid-window. Ahead at the end.

Every trade adds attribution data the agents learn from — so performance improves within the track, not between versions. Three phases, one direction:

100% = matching the market outperformance zone 01 · LEARNING 02 · CATCHING UP 03 · OUTPERFORMING 42% 78% 100% session 1 20 30 40 share of SPY's move captured 42% 78% 109% held positive while SPY dipped −3.5% final 10 sessions: +4.05% vs SPY +3.70%
Backtest, Apr 5 – May 31 2026 · rolling capture between labelled points is illustrative · not indicative of future results
PHASE 01 · LEARNING
Protected while it learned
Held positive through a −3.5% market dip, capturing 42% of the move.
PHASE 02 · CATCHING UP
Experience kicked in
Capture nearly doubled to 78% as accumulated decisions sharpened the calls.
PHASE 03 · OUTPERFORMING
Ahead of the market
Final 10 sessions: 109% capture — beating SPY outright, +4.05% vs +3.70%.

That arc — from 42% to 109% in ~40 sessions — is the point. Self-learning compresses a learning curve that takes human teams years into weeks, and the same compounding runs across every strategy on the platform.

The full window

Honest about the full window.

~40 SESSIONS · BACKTEST

AG-3 +9.05%

Apr 5 – May 31, 2026, cumulative return.

SAME WINDOW

SPY +15.70%

AG-3 trailed in absolute terms over the full window.

DRAWDOWN

−2.0% vs −3.5%

It did so at roughly half the drawdown of the benchmark.

We show the lag, not hide it. AG-3 traded return for control and improved as it learned. AG-3 v2 and a six-agent AG-4 are in development.

How the loop learns →