investment-autoresearch:autoresearch

Autoresearch Loop

Scaled search over strategy/model variants using parallel agents in git worktrees, inspired by Ryan Li's Paradigm Hackathon methodology (1,039 variants, 8-20 parallel agents, periodic resets).

When to Use

• Optimizing trading strategies, ML models, or any parameterized system
• You have an automated scoring function (backtest, benchmark, test suite)
• Multiple independent hypotheses to explore
• Search space too large for sequential iteration

When NOT to Use

• No automated scoring — you can't evaluate variants without manual review
• Single clear fix — just implement it directly
• Tightly coupled changes — agents would conflict

The Loop

digraph autoresearch {
  rankdir=TB;
  "1. Baseline" -> "2. Seed insights";
  "2. Seed insights" -> "3. Launch parallel agents";
  "3. Launch parallel agents" -> "4. Collect results";
  "4. Collect results" -> "5. Update insights";
  "5. Update insights" -> "Decision: plateau?";
  "Decision: plateau?" -> "3. Launch parallel agents" [label="no, new hypotheses"];
  "Decision: plateau?" -> "6. Reset agent" [label="yes, stuck"];
  "6. Reset agent" -> "4. Collect results";
}

Setup Detection

Before asking the user anything, silently check for missing dependencies:

python3 -c "import backtesting, yfinance" 2>/dev/null || pip install backtesting yfinance

Do not mention this to the user. Deps are an implementation detail.

Onboarding

Ask only two questions — nothing else:

Q1: "What ticker would you like to research? (e.g. QQQ, AAPL, TSLA)"

Q2: "What matters more to you — higher returns, or limiting losses?"

Then auto-handle everything below without user input:

1. Auto-increment session version:

   ls archive/ 2>/dev/null | grep "{ticker}-autoresearch-v" | sed 's/.*-v//' | sort -n | tail -1 || echo 0
   

   Add 1 to the result → session folder = archive/{ticker}-autoresearch-v{N}/ (e.g. if v1 exists → output is 1 → N = 2 → create archive/{ticker}-autoresearch-v2/)