Indian Government backed Soket AI Launches LOOP Harness for Next-Generation AI

0
108

Soket AI, a research lab building frontier AI models and systems that institutions can operate under their own control, unveiled LOOP, an open-source harness designed to enable AI agents to work continuously for weeks, not minutes.

The global frontier is no longer the chat window. Through 2025 and 2026, AI labs moved from models that answer, to agents that act, to harnesses that can stay on a task. OpenAI shipped that layer as Codex. Anthropic shipped it as Claude Code. LOOP is Soket AI’s entry in the same class. Open source. Bring any model. Built for work a team cannot hand to a hosted agent.

Agents are no longer limited to quick chats. They are running for hours, days, and weeks on real work. The longer an agent runs, the more it matters what it can touch and how complex a problem it can solve.

LOOP is a harness for that. It is lightweight and performant enough for the most critical workloads. It works for any developer and is designed for environments where trust is non-negotiable cybersecurity, banking, finance, and defence. Those are examples, not the limit.

The preview version brings together LOOP’s core agentic capabilities in a lightweight, high-performance architecture built in Rust. It is designed for low latency, efficient resource use, and parallel tool execution, making it easy to integrate into different environments. Developers can use any AI model by simply adding the model’s base URL, while LOOP automatically handles the rest.

LOOP also introduces flexible session management, allowing developers to branch, pause, and resume sessions, similar to how they work with Git. Session history is preserved during compaction, ensuring that context is not silently lost. A single task can also be distributed across multiple agents that share memory, with optional isolation through technologies ranging from rootless containers to microVMs.

Developers can turn skills and templates into simple slash commands, use MCP (Model Context Protocol), or serve their own MCP tools. Built in Rust, LOOP is designed to operate close to the system layer. Soket AI has developed the core technology and infrastructure in-house, building the agentic harness from the ground up.

In a local comparison using eight workers, with both systems pointed at the same Qwen model through an OpenRouter base URL and key, LOOP used 7.7 times less RAM and 17 times less local CPU than Claude Code. LOOP’s memory usage rose from 29 MiB to 223 MiB as workers were added, following a near-linear pattern. The measure is of the harness process only, not remote inference, and is not a quality score.

“A strong model cannot reach peak performance by itself,” said Abhishek Upperwal, Founder and CEO, Soket AI. “The longer it runs, the more the harness matters. LOOP is that harness, and it is open. Try it, break it, and tell us what you think.”

For workloads that cannot fail, Soket AI is building the next layer sandboxing, personas, multi-agent systems, layered memory, remote agents, trusted remote sandboxes, and group chat for agents. These are not part of this preview. Telemetry and a desktop app are also still in development.

LOOP is now available as a developer preview on GitHub for Linux, macOS, and Windows, giving developers a flexible open-source harness for AI systems they can fully control.