Show HN: Agent-relay – multi-agent workflows that learn from past runs (github.com)

The launch of Agent-relay introduces a novel approach to multi-agent workflows, enabling systems to learn from their past executions. This tool allows teams to capture and commit plans, review feedback, build logs, and audit trails directly into their repositories as markdown artifacts. Utilizing a feature termed "relay distill," the framework converts feedback, such as reviewer rejections, into actionable lessons that improve future task execution, allowing for a more efficient workflow and reducing time spent on repetitive tasks. Significantly, Agent-relay represents a paradigm shift away from traditional frameworks that rely on memory layers or hidden states. By maintaining all artifacts in a visible format within git, it promotes transparency and simplicity, allowing teams to edit, commit, and learn from their work collaboratively. Key technical features include its role-type lesson compilation, which ensures lessons are specific and relevant, and an integration with various backends for flexibility in implementation. This enhances productivity by responding dynamically to issues identified in earlier runs, ultimately making the planning and execution of tasks more effective and accurate within the AI/ML development landscape.