Artic: AI-Powered Multi-Agent Trading Platform

Deploy intelligent trading agents across any market — managed from a single hub.

1. Executive Summary

Artic abstracts the complexity of quantitative trading into a deployable, configurable agent. Where traditional bots require manual strategy selection and parameter tuning, Artic's LLM oversight layer continuously evaluates market regime, selects the optimal strategy from a library of 23 signal algorithms across 5 categories, and adapts on a configurable schedule — without human intervention.

The platform is designed around three core principles:

2. Problem Statement

The Gap Between AI and Execution

Algorithmic trading has long been the domain of institutional players with the engineering resources to build, maintain, and monitor complex systems. Retail traders and independent quants face a fragmented landscape: strategy research tools exist in isolation from execution infrastructure, LLM-based market analysis remains disconnected from live trading systems, and deploying multiple concurrent strategies requires managing sprawling, brittle codebases.

Specifically, the market lacks a platform that:

• Connects LLM intelligence directly to a live execution engine without manual intermediary steps.
• Enables concurrent multi-agent, multi-strategy deployment from a single control surface.
• Provides full observability — trade logs, LLM reasoning, position state — without requiring infrastructure expertise.
• Allows strategy research and live execution to share the same runtime environment.

Existing solutions — centralised copy-trading platforms, isolated strategy bots, and LLM-wrapped price feeds — address fragments of this problem. None orchestrate the full lifecycle from intelligent strategy selection through to live execution and monitoring at the agent level.

3. The Artic Solution

One Hub. Many Agents.

Artic introduces a hub-and-agent model where a central server (the Hub) maintains authority over all deployed agents. Each agent is an isolated Docker container — running locally or on a remote VM — with its own configuration, market scope, and strategy assignment. The Hub handles registration, lifecycle management, persistent state, and all client communication.

Users interact through whichever interface suits their workflow — a terminal UI, a web dashboard, a CLI, or a Telegram bot. All interfaces communicate with the Hub via a REST + WebSocket API. No interface has direct access to agents, ensuring a consistent, auditable control plane.

Communication between agents and the Hub is push-based. Agents POST status updates to the Hub every tick, push trade records on position open/close, and batch log entries every 10 ticks. The Hub never polls agents — all telemetry flows inward.

The LLM as Market Intelligence

At the core of each agent sits an LLM planning loop. On initialisation, the agent ingests historical OHLCV data and live price feeds, then prompts the configured LLM to assess market regime and select the most appropriate strategy from the active library. This evaluation recurs on a user-defined schedule — allowing the agent to shift strategy as market conditions evolve.

The LLM does not execute trades. Its role is advisory and supervisory: it reads structured market data, reasons about conditions, and outputs a strategy identifier with configuration parameters. The execution engine computes a signal via the selected strategy and acts on it within the risk parameters set at agent configuration time.

Every strategy in the library returns a normalised signal (a float between −1 and +1, where positive = bullish, negative = bearish, zero = neutral) paired with a diagnostic detail string. This uniform contract lets the engine treat all 23 algorithms identically regardless of their internal complexity.

Paper Trading & Safe Experimentation

Paper trading is the default mode. It uses identical strategy logic, LLM decision-making, and position tracking as live mode — but routes all order execution through a simulation layer. This allows users to validate strategy behaviour, observe LLM decision-making, and gain confidence before deploying real capital.

4. System Architecture

Artic is structured across six primary components. Each is independently deployable and communicates through well-defined interfaces:

Database

The Hub persists all state in PostgreSQL via SQLAlchemy async with Alembic migrations. The schema includes 9 models across core tables: users, agents, trades, log_entries, market_cache, user_secrets, agent_secret_overrides, and onchain_decisions. Log entries are append-only and bulk-inserted every 10 ticks for write efficiency.

Authentication

Artic uses a dual-method authenticationsystem. The Hub's auth dependency tries JWT first, then falls back to API key:

Agent Registry & Communication

The Hub maintains an agent registry — a live record of all running, paused, and completed agents. Communication is entirely push-based: agents push state to the Hub, the Hub never polls.

If an agent container crashes, trade history is preserved in full in PostgreSQL; only the live in-memory position state is lost.

5. Strategy Engine

23 Signal Algorithms

The strategy library is Artic's core quantitative layer. Algorithms are organised into five functional categories, each targeting a different market regime or signal type. The LLM planner selects among these based on its analysis of current conditions:

In addition to signal algorithms, the library exports 4 utility functions: kelly_size and vol_scaling_mult for risk sizing, plus session_filter and day_of_week_filter for time-based filtering — 27 total exports.

Signal Contract

Every strategy exposes a standard interface. It accepts price history and/or OHLCV candle data as input and returns a tuple:

The execution engine and the risk layer consume these outputs uniformly, regardless of which strategy is active. The dispatcher in signals.py routes a strategy name to the correct algorithm function.

LLM Strategy Selection Flow

6. Risk Management

Every Artic agent operates within a risk framework configured at deployment time via the risk_params JSONB column on the agents table.

Implemented Controls

7. Supported Markets & Integrations

Artic's integration layer is designed to be executor-agnostic. Any market or exchange that exposes a compatible order interface can be added as an executor module.

HashKey Global serves as the primary live execution target. Its perpetuals API provides order placement, position management, funding rate data, and balance queries — mapping directly to Artic's BaseExecutor interface. The sandbox environment is used for integration testing prior to mainnet deployment.

8. Product Roadmap

The Strategy Marketplace (Q3) transforms Artic from a personal trading tool into a collaborative ecosystem where quantitative researchers distribute and monetise their work. The multi-sector expansion (Q4) extends the addressable market beyond crypto derivatives into traditional finance instruments.

9. Risk Disclaimer