AWS lets developers write custom reward rules for multi-turn AI agents

AWS published guidance on building custom reward functions for multi-turn reinforcement learning in Amazon Nova Forge. Developers can now score an agent's behavior across an entire session rather than judging single responses, encoding their own success criteria directly into the training loop.

• Nova Forge supports reward functions as custom code in the training loop
• Scoring covers a full session, not a single reply
• Aimed at support and task-automation agents
• Available only to teams customizing models on AWS