跳到主要内容

Agent Studio

Agent Studio lets you create a trading agent with a character — no code — and give it management of your accounts toward measurable goals. An agent is like a personality-driven cBot: you pick an archetype and attitude, set the guardrails, and it runs under the Autonomy & Safety Kernel.

Open AI → Agent Studio (/agent-studio).

Create an agent​

The New agent dialog collects, no-code:

  • Name and archetype — Scalper, Day Trader, Swing Trader, Position Trader, News Trader, Contrarian, Mean Reversion or Breakout/Momentum. Each preset fixes a sensible cadence and posture.
  • Attitude — aggressiveness, patience and trend-following sliders.
  • Managed account(s) — at least one is required to create the agent (an agent with no account could never start, so Create stays disabled until you pick one). If you have not linked a trading account yet, the dialog says so and points you to link one first.
  • Autonomy level — Advisory (proposes only) or Approval-gated (acts only after your per-action approval). Full Auto (no per-trade approval) additionally requires a risk envelope and acceptance of the risk disclaimer before it can arm.

The persona compiles deterministically into the agent's system prompt (no LLM authors it), so the same configuration always produces the same instructions — reproducible and auditable.

The roster​

Every agent shows in a control-room table: which agent, its type, how many accounts it manages, its goals, run status, and last action, with Start / Stop / Kill controls. The Kill switch halts a running agent immediately.

Safety is a domain invariant, not a setting​

Everything money-touching routes through the Autonomy & Safety Kernel:

  • Risk envelope — hard per-order limits (max daily loss, open exposure, position size, leverage, consecutive losses, orders/hour, allowed symbols). Every order is validated against it before dispatch; a breach is refused, not clamped. Required before an agent can reach Full Auto.
  • Circuit breaker — deterministically halts new risk on a loss streak, a daily-loss breach, a hard performance-goal breach, or AI-provider unavailability (a down or hallucinating model never opens fresh positions).
  • Versioned disclaimer consent — a one-time, versioned acceptance is required to arm Full Auto (legally-required consent, not per-trade approval); bumping the disclaimer forces re-consent.
  • Kill switch — an idempotent emergency halt on every running agent.

Goals​

Give an agent measurable objectives — e.g. keep max drawdown below 4%, profit factor at least 1.5, win rate ≥ 55%. Each target is Hard (a guardrail — a breach trips the circuit breaker) or Soft (steers reasoning only), evaluated as On-track / At-risk / Breached.

The decision pipeline​

Once started, an agent runs a 24/7 supervised loop (AgentRuntimeService). Each tick, for every managed account, it: reads the deterministic account state (ground truth, never the model's memory); asks the decision engine for a move; passes it through the safety gate (AgentDecisionProcessor) — autonomy level → circuit breaker → risk envelope; writes an append-only AgentDecisionRecord; and halts or executes as the gate directs. The loop is fault-isolated (one agent's failure never touches another or the host) and safe by default: it is inert unless AI is configured and App:Ai:AgentRuntimeEnabled is set, and it never opens fresh risk while the AI provider is unavailable.

  • Approval gate — an Approval-gated agent's proposed order is recorded as Pending and does nothing until the owner approves it (POST /api/agent-studio/{id}/decisions/{seq}/approve or /reject); Full Auto clears through the envelope with no per-trade approval; Advisory only proposes.
  • Audit ledger — every decision is replayable: reasoning (XAI), the evidence it cited, the gate verdict, the order intent and whether it executed, at GET /api/agent-studio/{id}/decisions.
  • Research desk — an on-demand multi-agent debate: Alpha/Sentiment/Technical/Risk analysts each give a view and a Reviewer synthesises a proposal (POST /api/agent-studio/{id}/debate).
  • Memory — the agent remembers each decision and recalls recent memory into its next prompt for continuity (GET /api/agent-studio/{id}/memory).

Each roster row's Details opens the agent's decision feed (with Approve/Reject on pending orders), its memory, and a Run-debate tab.

Scope​

Shipped: the full agent lifecycle, the deterministic safety gate, the 24/7 runtime, the human-in-the-loop approval gate, the audit ledger, and the live cTrader Open API integration — the account-state store (reads real balance, positions and open exposure in lots) and the order executor (places real market orders, lots→volume via the symbol lot size), both resolving each managed account's OAuth credentials and degrading safely when an account is not linked. Requires the Anthropic API key for the model to generate orders (until then the engine holds); still to come are multi-agent debate roles and layered memory/reflection. The runtime is off unless App:Ai:AgentRuntimeEnabled is set, so live trading only happens on an explicit, fully-consented opt-in.

Managed accounts and editing​

When creating an agent you pick the trading account(s) it manages — at least one is required at creation (the Create button is disabled until one is selected, and the create endpoint rejects an empty selection). Every agent can be edited afterwards (name, temperament, autonomy, and managed accounts) from the pencil icon on its roster row. Lifecycle controls (details, edit, start, stop, kill) are icon buttons, each disabled in states where the action does not apply.