AI Trading Agents Hackathon - lablab.ai

AI Trading Agents
Harness

A harness for autonomous AI trading agents: multi-agent analysis, MCP trading platform, and a declarative risk engine with on-chain evidence via ERC-8004. Three pillars. Zero competitors.

Watch Demo View on GitHub
3x
Agent Architectures
45
MCP Tools
31
Risk Control Rules
4,772
Trades Registered On-Chain
Demo

Watch it in action

See how the platform handles live market conditions - agents analyzing, trading, getting blocked, learning.

📄 Download Pitch Deck (PDF)

Agents registered on-chain

Agent ID Name Stack Strategy Operator Wallet
32 Swiftward Alpha Go + Claude Code Momentum trader, regime-based 0x6Cd7...d13
Systematic momentum with discretionary overlay. Regime-based deployment (bull/neutral/bear). Claude Sonnet 4.6, 15-min interval. Lets winners run, cuts losers fast. Telegram streaming.
37 Random Trader Go Chaos baseline, 7 behavioral modes 0x7a2F...5e0
Scalp 40%, Swing 25%, FOMO 12%, Take Profit, YOLO, DCA, Panic Sell. Mood shifts every 15-25 ticks. 10-sec interval. The control group - proves the policy engine blocks bad trades.
43 Swiftward Gamma Go + Claude Code Multi-agent debate, 5 subagents 0xC5e0...F62
5 parallel analysts (technical, sentiment, market structure, bull/bear). 6 skills on demand: trading session, deep analysis, portfolio review, strategy update, morning brief, alert review. Claude Sonnet 4.6, 30-min interval. Telegram streaming.
49 Haia Trading Agent Python asyncio 3-stage deterministic quant 0xFa7b...0B7
Market health filter, asset rotation selector, decision engine with ATR stops and Kelly sizing. 4 loops (clock 15m, price spike 1m, tier2 5m, exit watchdog 2m). Pure math core - no hallucination where it matters.
77 Swiftward Midas Go + Claude Code Live Kraken, real capital 0xD890...b1A
Pullback trend trader on real Kraken account. Buys pullbacks in confirmed uptrends, targets 2R profit. Win rate >60% focus. Hard risk limits enforced by Swiftward policy. Claude Sonnet 4.6, 15-min interval.

Live servers - explore right now

Real infrastructure with historical trades and data. Random Trader active, AI agents require API keys to run.

📊 Trading Dashboard 🏟️ Backtesting Arena 📋 SigNoz Observability

Trading Dashboard: equity curves, positions, trade history, evidence hashes, policy decisions, agent activity.
Backtesting Arena: strategy runs, batch results, PnL comparison, agent selection for live trading.
SigNoz: logs, traces, and metrics across all services. Login: lablab@swiftward.dev / RfkxEvJ98w2ylpWgRw5!

Why We Win

Three Pillars.
Better in all three.

Every competitor built a trading bot. We built smarter agents, a trading platform, and a risk engine. Three pillars nobody else attempted.

🧠

Smarter Agents

Multi-agent debate. Deterministic quant trader. Backtesting arena. Self-improving with memory. Harness runtime. Code sandbox.

🔧

Trading Platform

6 MCP servers, 33+ tools. Multi-source market data. News, sentiment, Polymarket. Alerts and conditional orders. On-chain evidence via ERC-8004. Dashboard.

🛡️

Super Safe

Declarative YAML rules. Graduated risk tiers and circuit breakers. Safe rollout with backtesting and shadow mode. 3 gateways. Observability. Audit trail.

Pillar 1

Smarter Agents

Three agent architectures, code sandbox, self-improving memory, and the runtime that orchestrates it all.

🧠

Architecture 1: Multi-Agent Debate + Skills

5 subagents analyze every trade in parallel: technical, sentiment, market structure analysts + bull/bear researchers. 6 skills on demand: trading session, deep analysis, portfolio review, strategy update, morning brief, alert review.

🔬

Architecture 2: Deterministic Quant Trader

Python agent with 3-stage mathematical brain: market health filter, asset rotation selector, position sizing engine. Pure math for stops and sizing - no hallucination in the core layer. 4 concurrent loops.

📈

Architecture 3: Backtesting Arena

Multiple strategies evaluated via backtesting framework. Best-performing agent selected for live trading. Market simulator with MCP stubs for offline evaluation.

🧪

Python Code Sandbox

Isolated Docker container per agent. Downloads market data, writes and runs custom analysis scripts. Persistent state across sessions. The agent codes its own trading tools.

💡

Self-Improving

Persistent memory across sessions. Detailed session reports with trade reasoning. Self-reflection after losses - analyzes what went wrong, creates permanent rules. Strategy evolves from exploration to fine-tuned edge.

🛠️

Harness Runtime

Event-driven agent harness (think OpenClaw, but for trading and 100% isolated). Orchestrates Claude Code CLI via stream-json protocol. Session lifecycle, alert injection, Telegram streaming, error recovery. Works with API keys or Claude subscription for flat-cost operation.

News alert - false positive detection
News alert fires - agent checks sources, detects false positive, holds position
5 subagents spawned in parallel
3 analysts spawned, then Bull + Bear researchers debate
Agent self-analysis
Agent analyzes mistakes, creates permanent trading rules
Agent positions across 6 assets with concentration chart
Multi-asset portfolio: 6 concurrent positions with real-time PnL and concentration tracking
Arena batch results - cost, tokens, PnL stats
Batch #10 results: 6 runs, 1.7M tokens, best/worst PnL - pick the winner for live trading
Arena - 6 agent strategies evaluated
Backtesting arena: 6 strategies compete - Trend Follower, Quant Analyst, Portfolio Rebalancer, Mean Reversion, Macro, Scalper
Agent learns new tools and saves learnings
Agent discovers new MCP tools and saves session learnings to memory
Deterministic Python agent flow diagram
Deterministic agent: 4 triggers, cooldown gate, 3-stage brain, 5 data sources
Each agent gets isolated persistent folders
Each agent is fully isolated: personal Kraken CLI workspace and credentials, persistent workspace, and Claude Code home directory
Operator interacting with agent via Telegram
Telegram Integration

Operator in the loop.
Agent in control.

Agent streams output in real-time to Telegram. Operator can message mid-session - ask questions, request analysis, guide decisions. The agent acknowledges and incorporates.

  • Live output streaming as the agent thinks and acts
  • Two-way communication - ask agent to check an idea or trade
  • Risk engine alerts: halts, circuit breakers, blocked actions
  • Agent alerts: price triggers, news events, session summaries
Pillar 2

Trading Platform

Universal infrastructure. Any agent plugs in. Standardized MCP protocol.

🔧

6 MCP Servers, 33+ Tools

Trading (submit, estimate, portfolio, history, limits). Market Data (prices, candles, orderbook, funding, OI, alerts). Files, Code Sandbox, News, Risk. All JSON-RPC standard.

📊

Multi-Source Market Data

Kraken, Binance, Bybit connectors with degradation chain. PRISM API for asset resolution and signals. 7 server-side indicators (EMA, SMA, RSI, MACD, BBands, ATR, VWAP). Fear & Greed index. Funding rates. Open interest.

📰

News + Sentiment + Polymarket

CryptoPanic news feed with sentiment scoring. Event classification (hack, regulation, listing, macro). Polymarket prediction data. Keyword alerts that wake agents.

🔔

Alerts & Conditional Orders

Price alerts (above/below/change%). News keyword alerts. Auto-executing stop-loss and take-profit (OCO-linked). Soft stops that wake the agent without auto-selling. Trailing stops.

⛓️

On-Chain Evidence (ERC-8004)

We maintain our own keccak256 hash chain - signing all internal decisions AND external RiskRouter verdicts. Proofs posted to ERC-8004 Validation Registry. Agent identity as NFT. 6 reputation metrics on-chain.

🖥️

Professional Dashboard

6 screens: Overview, Agent Detail, Market, Evidence, Policy, Claude Agent. Equity curves, positions, trade history with evidence hashes. Single binary deployment.

Positions and asset concentration
Asset concentration with 50% policy limit
Open positions with profit
Open positions: FARTCOIN +15.6%, total unrealized +$271, SL/TP on all
Active alerts overview
5 active alerts - hard stops, take profits, soft stops with auto-execute
Alert triggered - MON breakdown
Soft stop triggers - agent analyzes breakdown, exits in 2 tranches
Active alerts dashboard - trading and market alerts
Platform alert monitor: trading alerts per order, market soft stops, and news sentiment watchers
Trades with evidence links
127 trades - each with decision hash, evidence link, reasoning
Evidence detail - full JSON proof with on-chain hash
Every evidence published to a public URL - full proof: trade reasoning, policy approval, on-chain tx hash
Decision hash chain - 339 decisions
339 decisions hash-chained - fills and rejects linked
Decision trace detail with JSON
Each decision: full trace JSON with reasoning, confidence, fill data
Pillar 3

Super Safe

Trading risk management and AI operations controls: graduated tiers, declarative rules, three gateways, full audit.

📝

Declarative Versioned Rules

Human-readable off-the-code policy rules. Change risk limits without redeploying. Every rule is versioned. Rulesets follow their own lifecycle: draft -> candidate -> active -> archived. Entity state per agent: counters, labels, buckets, metadata.

⚡

Trading Risk Rules Pack

Graduated risk tiers (limits tighten as drawdown grows). Circuit breakers (3 losses = 1h pause). Three-layer heartbeat drawdown. Stop-loss enforcement. Concentration caps. Pair whitelist. Rate limits. Operator halt/resume.

🧪

Safe Rollout: Evals, Backtesting, Shadow, A/B

Never deploy untested rules. Eval tests validate rules before rollout. Backtesting on historical events. Shadow mode runs next version of rules in parallel, logged not enforced. A/B testing compares policy versions on live traffic.

🔐

3 Gateways + Injection Defense

Agents are fully isolated - all communication goes through 3 gateways: MCP (tool calls), LLM (AI prompts), Internet (HTTP). Every request policy-evaluated. ML classifiers (PG2 + BERT) for prompt injection. LLM Gateway also enforces required behaviour - e.g. end-of-session attestation with reasoning for every agent.

📊

Observability & Scalability

OpenTelemetry metrics and logs (SigNoz). Kubernetes-ready. Horizontal scalability. Docker Compose for dev, multi-platform Docker images. Policy alerts via Telegram (halts, breakers, rejections).

🧾

Audit & Evidence Hash Chain

Every decision (approved or rejected) gets a keccak256 hash chained to the previous. Modify one record - the chain breaks. Public Evidence API. Full audit trail from genesis to latest trade.

Active policy rules
Declarative YAML rules - change limits without redeploying code. 16 rules, 301 rejections.
Trades list with policy rejection
FARTCOIN BUY rejected - 3 losses, 1h pause (exits allowed)
Policy analytics - 43K events, 668 violations, top rules
Policy analytics: 43,949 events processed, 668 violations (2%), events over time, top rules fired
Event investigator - trade_order events with verdicts
Event investigator: browse trade_order events - approved/rejected verdicts, matching rules, top entities
Loss streak protection
3 losses - automatic 1h trading pause
Blocked domain alert
Agent tried blocked domain - REJECTED by internet policy
Graduated tier block
Caution tier: $1K order blocked (>10%), agent reduces to $970, filled
SigNoz observability - full trade pipeline logs
SigNoz logs - two trades approved and filled, third blocked by policy. Full pipeline visible.
The Team

Built by Swiftward Team

Platform engineers building safe AI infrastructure for autonomous trading.

Konstantin Trunin

Founder & CTO, Swiftward

25+ years in distributed systems. Engineering leader across 15+ startups, Co-Founder & CTO of a logistics SaaS. Built the platform end-to-end.

Tikhon Slasten

Tech Lead, HAIA

6+ years in software engineering, focused on distributed systems and Web3 integrations. Builds scalable systems where AI meets on-chain finance.

Ivan Puchkov

Java | Kotlin Senior Backend Engineer

Senior Java/Kotlin Backend Developer with 5+ years of experience in designing and developing banking and financial systems

Ruslan Mukhametov

Ruby on Rails Lead Engineer

9+ years in FinTech. Built the backtesting and prompt eval layer — powered by SolidLoop, an open-source library for observable AI agents.

Oh, and one more thing...

Agent Intelligence

We audited every agent
in the hackathon

84 agents. 14,287 trades. $51M volume. On-chain attestations, reputation scores, sybil detection, AI-generated verdicts. On top of everything above - we built the intelligence layer to analyze every agent in the hackathon.

Agent Intel - hackathon agent leaderboard
84 agents. Scored, ranked, and audited. Explore Agent Intel →

Run it yourself

Open source. One command to start the full stack. Docker Compose, pre-built images, production-ready.

View on GitHub Watch Demo
git clone https://github.com/disciplinedware/swiftward-ai-trading-agents.git && cd swiftward-ai-trading-agents && cp .env.example .env && make up PROFILES=random