Castform and Neon are revolutionizing AI agent training by addressing the challenges of data usability and model efficiency. While traditional methods relied heavily on embedding search and handcrafted RAG pipelines, the emergence of agents introduced multi-hop search workflows. However, these complex searches often incurred significant costs and latency, with requests to frontier models like gpt-5.6-sol taking over 10 seconds and costing around $0.03.

Castform offers a solution by enabling developers to RL post-train open-source models, making them competitive with or even superior to frontier models at a fraction of the cost. This process involves turning existing company data into training tasks and managing the RL loop. Neon, with its Lakebase Search extensions on Postgres, provides the necessary infrastructure for efficient data retrieval and dynamic compute scaling, absorbing bursty workloads during training.

The integration allows Castform's training pipeline to leverage Neon's capabilities for synthetic data generation, RL training, and production inference. This partnership aims to make post-training as accessible as prompt engineering, empowering developers to build cheaper, faster, and more effective AI agents without deep ML expertise. The platform offers full observability into training runs, aiding in debugging and optimization.