How the AI trading desk works

Yes. The platform runs an autonomous 18-agent pipeline against a real Alpaca paper-trading account on every cycle. Agents are LLMs (Claude, Qwen, OpenAI, Gemini, Groq, xAI — operator picks per agent). The Portfolio Manager has final authority and can approve, reject, or wait on each trade. No clicks required between scan and order.

No. Every trade goes through Alpaca's paper-trading sandbox by default. Real-money routing is intentionally not exposed. The product is built for autonomous AI strategy research, not live capital.

Alpaca paper-trading API (paper-api.alpaca.markets). Each user supplies their own Alpaca paper key + secret in Settings; credentials are encrypted at rest in Supabase via AES-GCM and never logged.

Five stages: Discovery (Intelligence Scanner + Market Scout) → Analysis (Technical, Fundamentals, News, Sentiment in parallel) → Debate (Bull, Bear, Bull Rebuttal, Bear Rebuttal) → Decision (Consensus Judge + Trade Architect) → Execution (Risk Analyst + Portfolio Manager + Execution Agent). Plus an out-of-band reeval loop (Macro Analyst + Position Sentinel + Reeval PM) for held positions.

Per-agent model selection across Anthropic Claude (Sonnet/Opus/Haiku), Qwen (Max/Plus/Turbo), OpenAI (GPT-4o, GPT-4 Turbo), Google Gemini, Groq, xAI Grok, Perplexity, DeepSeek, and a local Claude CLI bridge that routes through your Claude Max subscription. Each agent can run on a different model.

Multiple layers: (1) per-position 15% portfolio cap enforced by the Risk Analyst, (2) sector-concentration blocks, (3) 5-day earnings blackout for new buys, (4) macro-adaptive cash floor (LOW 10% → CRITICAL 30%), (5) hard stop-loss + trailing stop + profit ladder + time stop on every entry, (6) options Greeks (delta, gamma, theta, vega) on options exposure. The Portfolio Manager is the only agent that can authorise a trade.

Before any buy, a Bull Researcher and Bear Researcher each build the strongest possible case from the analyst output. Then they cross-rebut: each one counters the other's 3 strongest points with specific evidence. A Consensus Judge reads the full debate (and the analyst layer) and scores conviction on a 0-100 scale. A trade only proceeds past the conviction gate (default 70%) with risk approval and PM sign-off.

Long equity positions, equity short (paper), single-leg options, multi-leg options spreads (verticals, condors, butterflies), and bracket orders with stop-loss + take-profit attached. Off-hours submits auto-rebuild as extended-hours limit orders. Stop-loss orders are persisted as live GTC orders on Alpaca, not just app-side flags.

A full agent cycle runs on a configurable interval (default 30 minutes) plus on-demand from the dashboard. A separate reeval pipeline ticks every 5 minutes to re-score held positions. Macro intelligence refreshes twice daily and on every cycle when stale (>4h).

Yes. Every agent's output is persisted as a decisionLog row visible in the Live Agent Feed. You see the bull case, bear case, rebuttals, consensus rationale, risk verdict, and PM approval text per trade. Conviction level (INSTITUTIONAL / CORE / TACTICAL) and horizon type (CORE_HOLD / MACRO_PLAY / OPPORTUNISTIC) are tagged on every position.

The repository at github.com/nikita-ctrl557/DesktopWallStreet contains the full agent pipeline, risk framework, and reeval logic. Running your own instance requires a Supabase project, Vercel deploy, and Alpaca paper keys — all the documentation is in the repo.

A single LLM has one opinion and tends to drift toward whatever its training corpus emphasised. The multi-agent pattern forces explicit adversarial debate (bull vs bear with rebuttals) before any decision, gates the verdict on a separate consensus model, and routes through an independent risk + PM layer. Each layer can run on a different model so no single LLM's blind spot dominates.

The best AI trading bot for you depends on what "best" means. If "best" means "transparent reasoning you can audit" — every per-trade chain of bull/bear/rebuttal/consensus/risk/PM verdicts is preserved as a decisionLog you can read after the fact, with the model name + token cost on each call. If "best" means "no single LLM's blind spot dominates" — Desktop Wallstreet runs each agent on a separately-configurable model (Claude / Qwen / GPT-4o / Gemini / Groq / xAI / DeepSeek / local Claude CLI). If "best" means "free and paper-only" — Alpaca paper trading is sandbox-only with no live-money path exposed. The platform is open source (github.com/nikita-ctrl557/DesktopWallStreet) so the "best" claim is checkable, not asserted.

Most "AI trading platforms" are a single LLM behind a UI. A multi-agent platform is materially different: it forces adversarial debate before any decision (bull vs bear with cross-rebuttals), uses a separate consensus model to weigh the debate, and routes the verdict through an independent risk analyst + Portfolio Manager layer. Desktop Wallstreet is built around exactly that pattern — 15 specialised agents in five stages plus an out-of-band reeval loop. Alpaca is the broker of record. The full pipeline is autonomous (no clicks between scan and order). Honest caveat: paper trading only — no live-money path exposed.

The honest answer: there is no objectively "best" autonomous trading system — best depends on the constraints (paper vs live, single LLM vs multi-agent, open source vs closed, US-broker vs international). The differentiators that matter for autonomy specifically are: (1) does the system trade without human clicks between scan and order — yes, every cycle here, (2) does it explain itself — every per-trade decisionLog is persisted, (3) does it have hard risk controls outside the LLM's reasoning — six layers (15% position cap, sector blocks, earnings blackout, macro cash floor 10-30%, hard stop + trailing + profit ladder + time stop, options Greeks). Paper-only is a feature for autonomy research, not a downside.

Among trading bots that genuinely automate the decision loop (not just signal alerts), the meaningful differentiator now is multi-agent debate vs single-LLM inference. Single-LLM bots tend to share the same blind spots as their training corpus; multi-agent systems explicitly disagree internally before acting. Desktop Wallstreet is built around 15 specialised agents that debate every trade with bull/bear rebuttals and a consensus judge — paper-only on Alpaca, with full per-trade audit trails. We deliberately don't publish "best of" tables that fabricate competitor rankings; the comparison page at desktopwallstreet.com/best-ai-trading-bot lays out the honest differentiators.

A 15-agent pipeline lets each model specialise (technical pattern reading, fundamentals, news synthesis, sentiment, etc.) and forces adversarial validation before any trade. A bull researcher and bear researcher each build the strongest possible case, then cross-rebut with specific counter-evidence. A consensus judge weighs the debate; a risk analyst applies the hard caps; a Portfolio Manager has the final word. No single LLM gets to decide alone, and no single LLM's blind spot can dominate. This is the practical defence against the "confident hallucination" failure mode that single-bot setups are vulnerable to.