Road map 2026 — Quant AI Agents

✓

Q1 2026 · Closed betaCompleted

Closed beta — measured the agents in the wild and locked the priors v1.0 ships with

Three months on live MT5 under real spreads gave us the data we needed: how often the Judge's confidence matched real-world R-multiples (reward-to-risk ratios), where Bull and Bear actually disagree, and the cost-vs-latency trade-off of running all five agents in parallel. Beta wasn't a feature launch — it was the controlled experiment that produced the priors baked into the Q2 binary.

✓Judge confidence calibration vs realized R-multiple — reliability diagrams + Brier-style scoring per regime

✓Bull↔Bear divergence distribution captured per regime; fed back into the executor as a meta-signal

✓Parallel-dispatch latency profile vs serial baseline; provider-failover recovery benchmarked at 99th pct

✓Token-budget envelope per decision profiled across major frontier providers under sustained load

✓Adversarial regime stress: news-flash + gap-open responses, prompt-drift detection on long-running sessions

✓Out-of-band ops surface: HMAC-signed Telegram rationale, chart-panel WebSocket bridge, replay-safe nonces

3 months under live spreads24+ instrumented testers1,200+ labeled AI Agents decisions

●

Q2 2026 · Public launchIn progress · Now

v1.0 — all five AI agents in a single MQL5 EA: deterministic, replayable, fully transparent

A bar-driven pipeline (Bull → Bear → Judge → Risk → Executor) wired into MT5 with every input exposed, deterministic prompt seeds, per-decision JSON logs, and a regime-gated execution path. Every threshold, model, prompt block and risk rule is an MT5 input. Nothing is hardcoded.

●5-agent deliberation pipeline (Bull/Bear debate → Judge confidence → Risk gate → Executor) with structured JSON contract between stages

●Multi-LLM router with provider fallback & budget cap across major frontier models; per-agent model assignment + temperature pinned for replay

●Regime classifier (trend / range / volatile) gates entry; vol-percentile + trend-strength bands; per-regime sizing & cooldown rules

●AI Hybrid SL: initial SL = max(structural swing, vol-anchored multiple), break-even reclaim on favorable extension, vol-aware trailing on continuation, time-stop on stale positions

●Optional position layering: capped DCA grid with hard kill-switch on regime flip, max-add count, and per-symbol exposure ceiling

●Per-bar decision log (JSON): inputs, prompt hash, agent outputs, judge score, executor action — full audit trail for replay & backtesting

5 agents · 1 bar tickFull input schema exposed<3s AI Agents round-trip

○

Q3 2026 · LabUp next · In design

Prompt Lab — a deterministic backtest where every prompt edit is scored against the equity curve

An in-browser simulator that replays MT5 bar history through the exact same pipeline as the live EA — same seeds, same prompt cache, same regime gate. Edit any prompt block and the lab tells you how much that change moved the equity curve, with confidence bounds from a bootstrap sampler. You see whether an improvement is real signal or just noise before it ever touches a real account.

○Bar-replay engine with tick-emulated fills, broker spread + commission + swap models, per-bar slippage distribution

○Walk-forward harness: rolling in-sample / out-of-sample split with re-fits, leakage-safe by construction

○Statistical layer: deflated Sharpe, Probabilistic Sharpe Ratio, Monte-Carlo bootstrap on trade-order to bound max-DD

○Pipeline determinism: pinned seeds + prompt cache + temperature 0 — identical run reproduces hash-for-hash

○Prompt diff with attributed equity delta — token-level highlight of which edit moved the curve, per regime

○Export: passing prompt → MT5 .txt + .set bundle, content-hashed for live audit trail

<2s per simulated barN=10k Monte-Carlo paths5y+ XAUUSD H4 loaded

♢

Q4 2026 · VisionFuture · Vision

Distilled AI Agents — your own GPU runs the model

By Q4, the v1.0 deployment will have produced a labeled dataset no general AI has: per-decision JSON capturing all five agent outputs, Judge scores, executor actions, and the real R-multiple that each trade hit. We distill that into a compact student model that runs on your machine and ends API rate-limit dependence on the hot path.

♢Corpus: per-bar JSON (5 agent rationales + judge score + executor action + realized R) across all v1.0 deployments

♢Distilled student: 3B–7B MoE encoder with multi-task heads (regime class, judge score, executor action)

♢Training: SFT on Judge outputs + outcome-weighted preference learning (RLAIF-style), walk-forward eval to block leakage

♢Deployment: GGUF Q4_K_M quantization, 8 GB consumer GPU, 50–200 ms inference per agent (target)

♢Hybrid routing: cloud AI Agents retained for high-conviction / high-R decisions, local model for filtering + low-stakes

♢Promotion gate: student must match cloud AI Agents on a held-out fraction of historical decisions before auto-routing

50–200ms local inferenceQ4_K_M @ 8 GB GPU0 API keys on hot path

An EA is a runtime.
The strategy is data.

— eatrading.ai · engineering principle

Determinism is non-negotiable

Every agent decision is replayable: pinned seeds, cached prompts, temperature 0, content-hashed inputs. A backtest that doesn't reproduce hash-for-hash is debt, not data.

The configuration is the API

Strategy lives in MT5 inputs, prompt blocks and .set files. The binary is the runtime; the configuration is the IP. Updates ship without a recompile, with an audit trail by file hash.

Cost has to scale to the user, not the lab

Token budgets, provider failover and on-device routing are first-class concerns. A model that wins on Sharpe but loses on dollars-per-decision doesn't ship.

Want to be in the next beta cohort?

Follow the Telegram channel and you'll see the call the moment we open spots.

Follow Telegram channel →