Configure which AI models run each agent role and how the multi-agent pipeline behaves. Most cost-vs-quality tradeoffs live here.
The five agents — Analyst, Strategist, Risk Officer, Compliance, and Decision — each take a model and an API key, plus six pipeline knobs at the end.
The AI that reads your chart and writes a structured technical report. It picks 3–8 of the most relevant indicators for the current regime,
measures them, and emits a JSON observation block consumed by the rest of the pipeline.
When to change
Upgrade to GPT-4o or Claude Sonnet for richer multi-indicator reasoning on choppy ranges; downgrade to GPT-4.1-nano if you re-analyze every 60s and cost is the dominant concern.
Provider API key for the Analyst agent. Accepts OpenAI sk-…, Anthropic sk-ant-…, or any OpenAI-compatible gateway key.
Stored encrypted in the terminal; never sent anywhere except the configured provider endpoint.
When to change
Paste once and reuse the same key in 1.4, 1.6, 1.8, 1.10 — that is the intended setup for solo users.
Watch out
All 5 agents share the same provider stack; if your key is rate-limited the entire pipeline stalls. Use a key with at least 5,000 RPM when re-analyzing every minute.
Synthesizes the Analyst's observations into a directional bias and a candidate entry plan (entry zone, invalidation, R-multiple target).
Does not place orders.
When to change
Upgrade only if you've already upgraded 1.1 Analyst — a stronger Strategist on a weak Analyst overfits to noise.
Same as 1.2 but for the Strategist. Reuse the same key 99% of the time.
When to change
Use a separate key only when you want per-agent billing dashboards or when running mixed providers (e.g. OpenAI Analyst, Anthropic Strategist).
Validates the Strategist's plan against position-sizing rules, current open exposure, and structural risk
(e.g. proximity to swing highs, asymmetric R:R). Vetoes plans whose effective R:R falls below 1.0.
When to change
The Risk Officer is the agent that most benefits from a stronger model on volatile assets — try GPT-4o on indices or crypto.
API key for the Risk Officer. Reuse the key from 1.2.
When to change
Provide a dedicated key only if you've enabled 2.11 Verbose log and want clean per-agent traces in your provider dashboard.
The macro/news filter. Cross-references the candidate trade window against high-impact events (NFP, FOMC, CPI, geopolitical headlines)
and blocks entries inside red-event windows.
When to change
Keep at default unless you specifically want a smaller model — Compliance runs less frequently than the others, so the savings are marginal.
Watch out
Compliance reasoning is only as fresh as the model's training/retrieval cutoff. For breaking news it's a filter, not a guarantee.
API key for the Compliance agent. Reuse the key from 1.2.
When to change
Same logic as the other agent keys — separate only for billing clarity.
The final aggregator. Takes the four upstream agents' outputs, weighs disagreement, and emits the binary
decision (OPEN_LONG / OPEN_SHORT / HOLD) plus a confidence score.
When to change
Highest-ROI quality bump on the page. Upgrading only Decision to GPT-4o typically improves trade-selection precision more than upgrading any other agent.
Watch out
The Decision agent is invoked once per re-analysis cycle. On 2.4 Re-analyze every = 60, that is 1,440 calls per day per chart — budget accordingly when upgrading the model.
API key for the Decision agent. Reuse the key from 1.2.
When to change
If you upgrade 1.9 to a more expensive model, this is the only key whose usage curve will change meaningfully — check it weekly.
Hard cap on how many indicator readings the Analyst is allowed to include in a single observation block.
Range 3–8. Higher values give the Strategist more context but cost more tokens.
When to change
Bump to 7 for ranging instruments where structure matters; drop to 3 for clean trending markets where one or two indicators dominate.
When true, the Analyst pulls bias from one higher timeframe (HTF = current × 6) before grading the entry timeframe.
Adds ~30% tokens per cycle but materially reduces counter-trend trades.
When to change
Turn off only on the highest timeframe you trade (no useful HTF above it) or when explicitly testing pure mean-reversion strategies.
When true, the Analyst additionally receives a rendered PNG of the current chart (200 bars, current symbol/TF) and the model is allowed to comment on visual structure (channels, double-tops, S/R clusters).
When to change
Useful with GPT-4o or Claude on visually-driven assets like indices or BTC. Negligible benefit on FX majors.
Watch out
Image inputs cost roughly 3–5× a comparable text call. If you enable this on 1.1 GPT-4o-mini the cost-per-cycle approximately doubles.
Controls the system-prompt budget given to each agent. Fast = terse one-shot; Standard = chain-of-thought with self-check;
Deep = multi-pass with explicit counter-argument.
When to change
Use Deep only on H4/D1 where you have minutes between cycles. Fast is fine on M5/M15 where speed matters more.
How long to reuse an Analyst observation if neither the price nor the bar index has changed materially.
Prevents redundant API calls during sideways minutes.
When to change
Set to 0 when debugging — every cycle becomes a fresh call. Set to 300 on D1 to silence redundant intra-bar calls.
Controls what the EA prints to the Experts log. Silent — only errors and trades. Compact — one-line per cycle with the Decision verdict.
Verbose — full per-agent reasoning (use sparingly; it floods the log).
When to change
Switch to Verbose when an entry felt wrong and you want to see exactly which agent voted which way.