Sarvam Launches Arya to Make Claude Code Built AI Agents Production Ready

Bengaluru-based Sarvam AI has introduced Arya, an agent orchestration stack designed to make LLM agents reliable enough for production workloads, after internal tests showed that conventional agent setups failed outright on a standard enterprise data extraction task.

In a 24 hour sprint, the company asked an agentic system to perform what it described as a routine ETL workflow. The goal was to extract about 200 structured finance and business metrics from 27 PDF reports of a listed company’s performance across multiple years. 

The first attempt, built as a single autonomous agent, delivered 0% accuracy. 

After six design iterations using Claude Code and custom development and debugging tools, the system improved enough to complete the workflow with acceptable accuracy.

Pratyush Kumar, co-founder of Sarvam AI, who shared the results publicly, said the team began with a straightforward specification and a single agent. It stalled within 15 to 20 minutes and attempted less than half the fields. 

“Arya system with GPT 4.1 mini achieved ~5x higher accuracy at ~10x lower cost compared to Claude Code with agent swarm,” Kumar said on X.

Logs showed repeated failures around units, reporting periods, and metric types. The team then widened the design to parallel agents and added structured orchestration, debugging, and deployment controls. Only then did the pipeline stabilise.

The takeaway, according to the team, was not about model intelligence but system structure.

The company argues that most current orchestration frameworks break in two places. They lack runtime reliability and they lack infrastructure grade guarantees around execution semantics, communication between agents, and persistent state. In practice, this leads to agents that fail silently and inconsistently, which makes them unsuitable for regulated or high stakes business tasks.

Arya is positioned as a middle layer between simple chains and heavyweight coding agents. At one end, lightweight LLM chains can run autonomously but handle only simple tasks. At the other, frontier coding agents can tackle complex problems but require heavy supervision and do not degrade gracefully. 

Arya aims to sit between these extremes by offering guardrails and repeatability without sacrificing flexibility.

Sarvam AI also highlighted the math behind reliability. Even with 99% reliability per step, a 50 step workflow succeeds only about 60% of the time. Adding infrastructure that enforces retries, state recovery, and deterministic execution changes that outcome, turning probabilistic behavior into something closer to enterprise software.

Arya will be open sourced. The release will include a containerised runtime that can run on any environment, a debug and replay interface to review runs, and an LLM assisted authoring layer. The company describes it as shared infrastructure, similar to databases or Kubernetes, rather than a proprietary tool.

For enterprises, the timing is relevant. Many are experimenting with agents for document processing, compliance reviews, and financial analysis. These workloads look simple but break quickly when agents hallucinate, skip steps, or lose state. Sarvam’s bet is that reliability, not raw capability, will decide adoption.

If the approach works outside controlled tests, Arya could become the plumbing layer that makes agent systems behave less like demos and more like software.

Sarvam AI has been on a 14-day launch spree. Over the last few days, the company has made several announcements like text-to-speech with Bulbul 3, speech-to-text with Saaras and vision models and ASR models that read and speak 22 Indian languages, even outperforming GPT and Gemini on several tasks.