githubinferredactive
autoresearch
provenance:github:darks0l/autoresearch
WHAT THIS AGENT DOES
Okay, here's a summary of the AutoResearch agent, geared towards a non-technical business user: What it does: AutoResearch is an AI system that automatically finds and improves trading strategies for decentralized exchanges (DEXs) on the Base blockchain. Think of it as a tireless researcher constantly experimenting with different ways to buy and sell digital assets. What real-world problem it solves: Most automated trading systems are stuck using the same, often outdated, strategies. When market conditions change, these systems fail. AutoResearch eliminates the need for human traders to constantly monitor and adjust these systems – it learns and adapts on its own. Who would use it: Investment firms, hedge funds, or even individual investors looking
README
<p align="center"> <img src="assets/darksol-banner.png" alt="DARKSOL" width="600" /> </p> <h3 align="center">Built by DARKSOL 🌑</h3> # AutoResearch — Autonomous DEX Strategy Discovery > **Karpathy-style autoresearch for Base DEX trading** — an AI agent that iteratively mutates, backtests, and evolves strategies against real Uniswap V3 + Aerodrome data on Base. Every experiment is remembered via LCM (Lossless Context Management), making each mutation smarter than the last. Built for [The Synthesis Hackathon](https://synthesis.md) (March 13–22, 2026). [](LICENSE) [](https://nodejs.org) [](test/) [](https://base.org) [](https://bankr.bot) [](https://openclaw.ai) --- ## The Thesis Most trading bots are static. Someone writes a strategy, deploys it, and prays. When markets shift, the strategy breaks and a human has to manually intervene. **AutoResearch eliminates the human bottleneck.** An AI agent runs a continuous research loop — proposing hypotheses, testing them against real Base DEX data, keeping only improvements, and learning from every failure. The agent doesn't just execute trades; it **discovers** how to trade. The key insight: **LCM memory makes the agent learn from its own research history.** Instead of blind mutations, the agent queries what parameter ranges work, what signal combinations improve Sharpe, and what approaches consistently fail. Each experiment is smarter because the agent remembers every previous one. ## Built in 12 Hours This entire system — 14 source modules, 51 tests, 230+ experiments, 71+ live trades, x402 payments, regime detection, daemon service — was built from **zero to production in a single 12-hour session** during The Synthesis Hackathon (March 21-22, 2026). **Daemon continues running post-session, autonomously iterating.** **Timeline:** - **Hour 0-2:** Core engine (indicators, backtest, scoring, memory) - **Hour 2-4:** Benchmark strategies + autonomous research loop - **Hour 4-6:** Bankr LLM Gateway integration — first 75 LLM-driven experiments (score: 0.421 → 0.740) - **Hour 6-8:** Real data pipeline, regime detection, production execution engine - **Hour 8-10:** Daemon service, live Bankr trades on Base, x402 micropayment service - **Hour 10-12:** Score breakthrough (2.838 → 8.176), 70+ live trades, Devfolio submission, Moltbook posts No pre-existing codebase. No templates. Built from `npm init` to autonomous daemon discovering strategies and executing trades on Base mainnet. ## How It Works ``` ┌─────────────────────────────────────────────────────┐ │ AutoResearch Loop │ │ │ │ 1. Read strategy.js + full score history │ │ 2. Query LCM: "what worked? what failed?" │ │ 3. LLM proposes ONE targeted mutation │ │ 4. Backtest against 4 Base DEX pairs (real data) │ │ 5. Score improved? → KEEP (commit + log) │ │ Score worse? → REVERT (log failure reason) │ │ 6. Update memory → repeat (smarter each time) │ │ │ │ Every experiment logged. Nothing lost. Agent learns. │ └─────────────────────────────────────────────────────┘ ``` ## Architecture ``` ┌────────────────────────────────────────────────────────────────┐ │ AUTORESEARCH ENGINE │ │ │ │ ┌─────────────┐ ┌─────────────────┐ ┌──────────────────┐ │ │ │ Controller │→│ Strategy File │→│ Backtest Engine │ │ │ │ (mutate → │ │ (single source │ │ (fee model, │ │ │ │ test → │ │ of truth) │ │ scoring, DD) │ │ │ │ learn) │ │ │ │ │ │ │ └──────┬───────┘ └─────────────────┘ └────────┬──────────┘ │ │ │ │ │ │ ▼ ▼ │ │ ┌─────────────┐ ┌─────────────────┐ ┌──────────────────┐ │ │ │ LCM Memory │ │ 10 Indicators │ │ Reporter │ │ │ │ (experiment │ │ RSI, MACD, BB, │ │ (batch reports, │ │ │ │ history, │ │ ATR, VWAP, │ │ Discord/TG) │ │ │ │ patterns) │ │ Stochastic... │ │ │ │ │ └──────────────┘ └─────────────────┘ └──────────────────┘ │ │ │ │ ┌─────────────────────────────────────────────────────────┐ │ │ │ BANKR INTEGRATION │ │ │ │ LLM Gateway (mutations) │ Wallet (live execution) │ │ │ │ Balance checks │ Trade execution │ Portfolio tracking │ │ │ └─────────────────────────────────────────────────────────┘ │ │ │ │ ┌─────────────────────────────────────────────────────────┐ │ │ │ DATA LAYER (Base DEX) │ │ │ │ ETH/USDC (Uni V3 0.05%) │ ETH/USDC (Uni V3 0.3%) │ │ │ │ cbETH/WETH (Uni V3) │ AERO/USDC (Aerodrome) │ │ │ │ 500+ hourly candles per pair │ Real on-chain data │ │ │ └─────────────────────────────────────────────────────────┘ │ └────────────────────────────────────────────────────────────────┘ ``` ## Quick Start ```bash git clone https://github.com/darks0l/autoresearch.git cd autoresearch # Fetch historical Base DEX data node scripts/fetch-data.js # Run benchmark strategies node scripts/run-benchmarks.js # Launch autonomous research (30 experiments) node scripts/run-autoresearch.js --max 30 # Run persistent daemon (auto-iterates, auto-commits) node scripts/daemon.js --batch 15 --model claude-sonnet-4.5 --target 10.0 # Run tests (45/45) npm test ``` ## Modules (14 Source Files) | Module | File | Purpose | |--------|------|---------| | **Controller** | `src/controller.js` | AutoResearch loop — mutate → backtest → keep/revert → learn | | **Backtest Engine** | `src/backtest.js` | Full backtester with fee model, composite scoring, drawdown tracking | | **Indicators** | `src/indicators.js` | 10 pure-math indicators (RSI, MACD, BB, ATR, VWAP, Stochastic, OBV, EMA, Williams %R, Percentile Rank) | | **Regime Detection** | `src/regime.js` | Market regime classifier — Hurst exponent, trend strength, volatility ranking | | **LCM Memory** | `src/memory.js` | Experiment logging, pattern queries, session-persistent learning | | **Data Layer** | `src/data.js` | Historical data loading for 4 Base DEX pairs | | **Data Feed** | `src/datafeed.js` | Production multi-source data pipeline (DeFiLlama + CoinGecko + Base RPC) | | **Execution Engine** | `src/executor.js` | Live trading via Bankr — risk management, position sizing, paper/live modes | | **Bankr Integration** | `src/bankr.js` | Bankr LLM Gateway for mutations + wallet for live execution | | **Reporter** | `src/reporter.js` | Batch reports, final summaries, Discord/Telegram formatting | | **Pair Discovery** | `src/discovery.js` | Manual pair add/remove + auto-scan top Base DEX pools by TVL | | **Config** | `src/config.js` | Centralized configuration, model selection, thresholds | | **Daemon** | `scripts/daemon.js` | Persistent autonomous runner — batch experiments, auto-sync, credit tracking | | **Report Generator** | `scripts/report.js` | Human-readable reports (Markdown + styled HTML) | | **Entry Point** | `src/index.js` | Public API — all exports for skill/library usage | ## Scoring System Each strategy is scored with a composite metric designed to re [truncated…]
PUBLIC HISTORY
First discoveredMar 22, 2026
IDENTITY
inferred
Identity inferred from code signals. No PROVENANCE.yml found.
Is this yours? Claim it →METADATA
platformgithub
first seenMar 21, 2026
last updatedMar 21, 2026
last crawledtoday
version—
README BADGE
Add to your README:
