Sarvam launches Arya to make AI agents production-ready in India’s sovereign AI push

On February 10, Indian AI startup Sarvam AI introduced Arya, a groundbreaking agent orchestration platform aimed at enhancing the reliability of autonomous AI systems for practical applications. This development is part of Sarvam's broader initiative to create sovereign AI tools tailored for the Indian market. The launch has garnered significant attention in the tech industry, particularly regarding the potential of "agentic AI"—models capable of autonomously planning and executing tasks. However, there is increasing frustration over the inconsistencies these systems exhibit when implemented at scale. As highlighted in Sarvam's official blog, while advanced models can perform complex tasks like writing compilers, they often fail to operate smoothly in real-world scenarios. "These agents do not fail gracefully. Instead, they fail silently and unpredictably, leading to costly errors," the company noted. In practical tests, Sarvam evaluated frontier AI agents on a standard enterprise task: analyzing financial reports from publicly traded companies to extract approximately 200 structured data points, such as revenue and risk disclosures. The company described this process as a common ETL (extract, transform, load) workload. Despite the simplicity of the task, Sarvam found that single-agent systems frequently encountered issues, such as misinterpreting financial units or failing to verify reporting periods. Even multi-agent approaches struggled to deliver consistent results. Sarvam emphasized that the obstacles faced were not due to a lack of intelligence but rather the framework's structure. "We believe these challenges stem from an infrastructure problem rather than a gap in model capabilities," the company stated. Arya is designed to address these issues by providing a framework that ensures reliability. It is built on four core principles: composable primitives, immutable state, controlled dynamism, and declarative authoring. Essentially, Arya deconstructs AI systems into manageable components, imposes strict data handling rules, and utilizes configuration files for workflow definitions instead of custom coding. This approach aims to enhance the effectiveness of AI applications in production environments. The unveiling of Arya follows a series of recent product launches by Sarvam, including Sarvam Dub, Sarvam Audio, Sarvam Vision, and Bulbul V3, positioning the company as a significant player in India's sovereign AI landscape. According to Sarvam, the motivation behind developing Arya was a response to the inadequacies of existing agent frameworks in production settings. "We created Arya because we had no alternatives; nothing available could withstand real-world application." As AI capabilities advance, Sarvam argues that the demand for resilient infrastructure will only increase, stating, "Models are akin to compilers, while infrastructure serves as the runtime. You wouldn't deploy software without a robust operating system. The future of agent technology will hinge not on the most advanced models but on providing developers with the infrastructure necessary to ensure reliability."