Evolution

The evolution subsystem optimizes agent system prompts using APO (Automatic Prompt Optimization) during GPU idle periods. It generates candidate prompts, evaluates them against held-out tasks, and promotes winners.

Evolution Cycle

1. Wait for GPU idle (<30% utilization)
2. Select next agent (round-robin)
3. Generate candidate prompts
4. Evaluate against held-out tasks
5. Promote best if improved
6. Record lineage

Optimization Methods

Model Merging

Agent versions can be combined using parameter-space merging:

Agent Versioning

Each evolution cycle produces a new agent version with tracked lineage:

# Check evolution status
uv run gaius-cli --cmd "/evolve status" --format json

# View agent versions
uv run gaius-cli --cmd "/evolve versions leader" --format json

# Promote a specific version
uv run gaius-cli --cmd "/evolve promote leader v3" --format json

Configuration

evolution {
    enabled = true
    idle_threshold = 60    # seconds of GPU idle before triggering
    cycle_interval = 3600  # minimum seconds between cycles
}

The daemon runs in the engine process and activates only during GPU idle periods to avoid competing with interactive inference.