Autoresearch

Karpathy-style keep-or-discard loop for runtime strategy self-improvement.

How It Works

Agent runs for N ticks
Evaluates performance: win rate, P&L, drawdown, avg win/loss
LLM proposes parameter changes based on results
Safety bounds are checked
Human approves or rejects

Self-Evaluation


forge strategy view ./my-agent


version: 1
parameters:
  spread_pct: 1.0
  position_size_pct: 1.0
  momentum_threshold: 0.5

performance:
  Win rate:     62% (target: >45%)
  Max drawdown: 3.2% (target: <5%)
  Avg win/loss: 1.8 (target: >1.5)

Mutation Proposals

The LLM analyzes performance and proposes changes:


IMPROVEMENT PROPOSAL [prop_a1b2c3]

Hypothesis: Relax entry confirmations
Change: momentum_threshold 0.5 → 0.3
Expected: +15% more trades

Safety check:
  position_size ≤ 25% ✓
  stop_loss ≤ 20% ✓

Safety Bounds

The agent cannot propose unsafe parameters:

Parameter	Max allowed
`position_size`	25%
`stop_loss`	20%
`leverage`	5x
`max_daily_trades`	100

Accept or Reject


# Accept the proposal — strategy advances to v2
forge strategy accept prop_a1b2c3
 
# Reject — keep current version
forge strategy reject prop_a1b2c3

Enable


forge run ./my-agent --autoresearch --eval-interval 6

--eval-interval controls how many ticks between self-evaluations (default: 10).