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

Homegrown artificial intelligence (AI) startup Sarvam AI on February 10 launched Sarvam Arya, an agent orchestration stack designed to make autonomous AI systems reliable enough for real-world production use, as the company accelerates its race to build sovereign AI tools for India.

The launch comes amid growing industry excitement around “agentic AI” systems, AI models that can plan, act and coordinate tasks autonomously, but also mounting frustration over their instability when deployed at scale.

“Frontier models can write compilers and rebuild browser engines, but when you try to run these agents on production workloads, they don’t fail gracefully,” Sarvam said in a blog post announcing Arya. “They fail silently, expensively, and differently every time.”

Sarvam said it tested frontier AI agents on a common enterprise task: analysing financial reports of publicly listed companies to extract around 200 structured data fields, including revenue, capital expenditure, shareholding patterns and risk disclosures.

“This is ETL,” the company said, referring to extract, transform, load, a standard data engineering workload. “The most well-understood workload in software.”

Despite this, Sarvam said a single-agent approach often stopped midway, misread units such as rupees versus percentages, or failed to confirm reporting periods. Even multi-agent systems struggled to complete the task reliably.

“Same model. Same task,” the company said. “The difference wasn’t intelligence, it was structure.”

Also read: Sarvam AI partners with Odisha and Tamil Nadu to build national compute grid

Infrastructure, not intelligence

Sarvam claims that Arya is an attempt to solve what it sees as an infrastructure problem rather than a model-capability gap.

“A framework is opinions about code. Infrastructure is guarantees about execution,” Sarvam said.

The company said Arya is built around four guarantees: composable primitives, immutable state, controlled dynamism and declarative authoring.

In simple terms, Arya breaks AI systems into small building blocks, enforces strict rules on how data is written and stored, limits where AI models are allowed to think freely and uses configuration files rather than custom code to define workflows.

Sarvam’s sovereign AI push

The Arya launch follows a series of product releases by Sarvam since early February, including Sarvam Dub, Sarvam Audio, Sarvam Vision, and Bulbul V3, as the company positions itself as a key player in India’s sovereign AI ecosystem.

Sarvam said it built Arya after repeated failures with existing agent frameworks in production.

“We built Arya because we had no choice,” the company said. “Nothing available survived contact with production.”

The company argued that as AI models become more powerful, the need for robust infrastructure will only grow.

“Models are compilers. Infrastructure is the runtime,” Sarvam said. “You wouldn’t ship a binary without an operating system.”

“The decade of agents won’t be defined by who builds the most capable model,” it added. “It’ll be defined by who gives developers the infrastructure to make that capability reliable.”