>_TheQuery
// Reading nowStart
← All Articles

Sarvam Code Launched With a New Three-Agent Architecture. Every Coding Agent Already Has One.

By Addy · July 29, 2026

Sarvam unveiled Sarvam Code at its Epoch 2026 conference, an India-hosted coding agent split into three specialized roles: a planner, a worker, and a verifier. Cofounder Pratyush Kumar framed the architecture as a fix for what he called one of open source's biggest weaknesses, reliably knowing when a coding task is actually finished. Using Terminal-Bench 2.1, Sarvam says its agent spent about 2persolvedtaskagainst2 per solved task against 4.1 to $27.8 for Claude Code and Codex, and that it beat both on the same benchmark. Kumar's pitch to Indian IT services firms was direct: stop paying Anthropic for coding licenses, this is hosted here, and it's competitive.

The problem Kumar named is real. Knowing when an agent has actually finished a task, rather than just stopped, is a genuine open question in this space. What's less clear is that splitting a coding agent into a planner, a worker, and a verifier is a new answer to it.

Plan, Act, Observe, With Job Titles

Every serious coding agent on the market already runs some version of the same loop: plan the change, act on it by editing files and running commands, observe the result, repeat until done. That's the documented shape of Claude Code, Codex, and OpenCode alike. Formalizing that loop into three named roles, calling one of them a verifier, is an implementation choice, not a departure from how agentic coding already works.

It's also not even a new implementation choice. OpenCode, the open-source, model-agnostic coding harness that anyone can already run today, supports running multiple agent instances in parallel on the same project specifically so one can draft an implementation while another reviews or tests it, using whatever models you choose for each role. That is functionally the planner-worker-verifier split Sarvam is pitching as new, available today, open source, for free, to anyone willing to configure it. Sarvam built a specific, fixed version of a pattern the open-source community had already generalized.

The Sovereignty Pitch That Isn't the Most Sovereign Option

Sarvam's other headline claim is data control: hosted in India, not routed through a foreign cloud. That's a real and legitimate selling point for Indian enterprises with data residency requirements. It is also not the most sovereign option currently available to those same enterprises.

OpenCode already supports fully local execution through Ollama, meaning a company that actually wants its code never leaving its own infrastructure can already do that today, with no dependency on Sarvam's servers, Anthropic's servers, or anyone else's, regardless of which country they sit in. Hosted in India is a real improvement over hosted in the US for a company worried about jurisdiction. It is still hosted somewhere that isn't the customer's own machine. If data sovereignty is the actual pitch, the open-source alternative already clears that bar more completely than Sarvam Code does.

A Vendor's Own Math, Not Yet Anyone Else's

The 2−per−taskfigureagainstClaudeCodeandCodex′s2-per-task figure against Claude Code and Codex's 4.1 to $27.8 is Sarvam's own benchmark, run and reported by Sarvam, on a product that just entered early beta. No independent evaluator has reproduced it yet, and Sarvam's own disclosure that "retries and dead ends are not billed" is the kind of billing detail that sounds generous and is also difficult to audit from outside the company running it. None of that makes the number false. It makes it exactly the kind of vendor-reported comparison that has needed a second opinion every other time a lab has published one this year, and this one hasn't gotten one yet.

What actually is distinctive about Sarvam has nothing to do with the agent loop. Sarvam-105B and Sarvam-30B were trained from scratch on Indian compute, built specifically for Indian languages and code-mixed text, and open-sourced back in February. That is the genuinely hard, genuinely sovereign part of what this company has built. Wrapping it in a planner, a worker, and a verifier, then marketing the wrapper as the breakthrough, undersells the one part of Sarvam Code that competitors actually can't just replicate with an open-source harness and a weekend.

Sources:

Previously on TheQuery: