Soket AI has launched LOOP, an open-source agentic harness designed to enable artificial intelligence agents to work continuously on tasks for extended periods, as developers increasingly experiment with autonomous systems that operate beyond short chatbot interactions.
The IndiaAI Mission startup has released LOOP as a developer preview for Linux, macOS and Windows. The platform is designed for AI agents that may need to operate for hours, days or weeks while using multiple tools and retaining context across longer workflows.
Built in Rust, LOOP is designed for low-latency operation, efficient resource utilisation and parallel tool execution. Developers can connect different AI models to the harness by providing a model's base URL, allowing the system to work across models rather than being tied to a single provider.
The platform also introduces session management capabilities that allow developers to branch, pause and resume agent sessions. Session history can be preserved during context compaction, while a task can be distributed among multiple AI agents that share memory.
For workloads requiring greater separation, LOOP supports optional isolation through environments ranging from rootless containers to micro virtual machines. It also supports the Model Context Protocol, or MCP, and allows developers to convert skills and templates into slash commands or serve their own MCP tools.
Soket AI is positioning the open-source system for applications including cybersecurity, banking, finance and defence, where organisations may require greater control over the infrastructure, data and tools accessed by autonomous agents.
The company has also published initial performance comparisons for the harness. In a local test involving eight workers, with LOOP and Claude Code connected to the same Qwen model through OpenRouter, Soket AI said LOOP consumed 7.7 times less RAM and 17 times less local CPU.
The comparison measured the harness process and excluded remote model inference. The company also clarified that the figures do not represent a comparison of model or output quality. LOOP's memory consumption reportedly increased from 29 MiB to 223 MiB as workers were added.
Soket AI Founder and CEO Abhishek Upperwal said that as models operate for longer periods, the infrastructure surrounding them becomes increasingly important to their performance.
The company is developing additional features including sandboxing, personas, multi-agent systems, layered memory, remote agents, trusted remote sandboxes and group chat for agents. These capabilities are not part of the current preview.
LOOP forms part of Soket AI's broader work on AI infrastructure that organisations can operate under their own control, alongside its foundation model and inference initiatives.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.