Sarvam AI launches multi-agent orchestration platform

AI agents are advancing at a remarkable pace. Yet when deployed, they still fail quietly, cost real money, and break in unexpected ways. This significant gap between potential and performance is what Sarvam AI aims to bridge with its new orchestration stack, Arya.

Traditionally, developers have treated AI agents like simple scripts, yet when these agents are required to handle hundreds of documents, they often collapse under the weight of their own context.

Sarvam AI co-founder Pratyush Kumar explains that the current landscape is split between simple chains that handle straightforward tasks and complex coding agents that need constant human supervision.

“The missing middle is production-grade agents - systems that are genuinely sophisticated but run at scale with reliability, thousands of times a day, without a human in the loop,” Kumar wrote in a post on X.

Arya introduces a structured approach to building these systems through four main guarantees. First, it uses eight flat primitives, which are basic building blocks such as Ledgers for data and Task Graphs for workflows. This ensures that capability emerges through how pieces are connected rather than just adding more layers of complexity. 

“Infrastructure doesn’t add reliability, it changes the math. Get all four right and reliability scales with complexity instead of collapsing under it,” noted a blog post on Arya.

One of the most significant shifts is how Arya manages information through an immutable state ledger. In most frameworks, agents can overwrite data like someone scribbling on a shared whiteboard, which leads to corruption if the system crashes. Arya treats the state like an accounting ledger where every entry is append-only and cannot be changed.

This means that if a process fails, the system can simply restart from a clean checkpoint without the “crime scene” of half-updated data.

The benefits of this approach were highlighted in a test involving the extraction of 200 financial metrics from 27 company documents. This is a task often called ETL, which stands for extracting, transforming, and loading data into a usable format.

Initial attempts using standard models alone struggled with units and reporting periods. However, after several design iterations using Arya, the system reached 86% accuracy. This was achieved at a cost of $1.20 per run, which is 10x cheaper than using agents without this specialised infrastructure.

As AI models become more capable, the need for a robust runtime or operating system to manage them becomes even more vital. Just as faster processors in the past required better operating systems, more intelligent AI requires better orchestration to remain useful in a business environment.

Kumar believes “more capable models amplify the value of a well-structured system, and well-structured systems give models more to work with”.

By providing a stable layer for execution, Arya suggests that the next decade of AI will be defined by reliability rather than just raw intelligence.

Arya’s launch follows a rapid series of announcements from Sarvam over the 1o days, including sovereign AI partnerships with multiple states; Bulbul V3, a text-to-speech model for natural, production-ready Indian-language voices; Sarvam Vision; Sarvam Audio, an audio extension of the 3B-parameter Sarvam language model trained on English and 22 Indian languages; and Sarvam Dub, an AI dubbing model that helps creators extend the reach of their content.