What problem does it solve?
This skill solves the problem of inefficient and centralized job distribution in a fleet of autonomous agents by enabling a deterministic, reputation-based auction system that selects the most cost-effective worker for any given task.
Core Features & Use Cases
- Deterministic Auctioning: Uses Thompson sampling to resolve job awards without a central auctioneer, ensuring reproducibility across all hosts.
- Reputation-Based Selection: Automatically tracks worker performance and cost metrics to favor efficient, high-quality implementations over simple token-cost metrics.
- Use Case: When a new coding task is posted to the garden, this skill allows multiple worker agents to bid on the job based on their historical success rate and estimated cost, ensuring the most capable agent handles the work.
Quick Start
Enable the bid-auction skill by setting the environment variable GARDEN_QUOTA_ROUTING to auction in your worker configuration.