Article complete

Soket AI Launches Loop, Open-Source Harness for Long-Running AI Agents

The company is training a 26-billion-parameter model that it plans to release as open source, along with its dataset and training stack.

A man in a blue suit gestures while presenting on stage, with the Soket logo and flowchart graphics in the background.

Key takeaways

  • Soket AI has launched Loop, an open-source harness for long-running AI agents, designed for efficiency and low resource usage.
  • The system allows developers to run multiple agents simultaneously, with features like parallel tool execution and memory sharing.
  • Loop supports the Model Context Protocol and enables the conversion of skills into slash commands for better usability.
  • Initial benchmarks show Loop's performance is competitive, with low median task times and resource consumption compared to other harnesses.
  • The platform is aimed at applications in areas such as cybersecurity and defence, showcasing its versatility in real-world tasks.

Takeaways drafted with AI from the article text.

“Agents are no longer a quick chat window. They're running for hours, days, weeks on real work,” Upperwal said.

Soket AI said Loop is designed to run multiple agents efficiently, with parallel tool execution and low CPU and memory requirements. It is built on Rust, which the company said helps keep the system lightweight and suitable for running agents locally.

Developers can connect different models through a base URL, while sessions can be forked and operated like Git.

The company also said a single goal can be split into a swarm of agents that share memory, with isolation options ranging from rootless containers to micro virtual machines. Loop also supports the Model Context Protocol (MCP), while skills and templates can be converted into slash commands.

In its initial benchmarking, Soket AI said Loop's total runtime was close to Pi, a minimalist agent harness, while its cost performance was ahead of OpenCode and Goose. It noted that Loop recorded the lowest median task time among the harnesses it compared, along with low peak memory and CPU use.

Upperwal said these characteristics could allow thousands of agents to run on a small system. A demonstration at Cypher showed Loop researching the day's conference schedule through multiple tools, before returning a list of talks and sessions it considered relevant.

Soket AI said Loop is being developed for cybersecurity, defence, banking, finance and legal workloads. The company said sovereign deployments in these sectors require capabilities such as air-gapped environments, zero-trust execution, auditability and support for long-running tasks.

The roadmap includes sandboxing personas, multi-agent systems, leveled memory, remote agents, trusted remote sandboxes and group chat for agents. Upperwal said these workloads require a high degree of determinism, particularly in defence.

The Loop launch is part of Soket AI's broader focus on Math Code Reasoning (MCR) agents. Upperwal said the company has aggregated close to 30 trillion training tokens, primarily focused on mathematics and code, along with nearly two billion reasoning traces.

Agents are no longer a quick chat window. They're running for hours, days, weeks

Upperwal

The company said its MCR approach combines mathematical reasoning, natural-language reasoning and code execution to enable agents to interact with their environments and perform longer-running tasks.

Upperwal said Soket AI open-sourced Loop to encourage wider adoption and community contributions. He also announced that the company is training a 26-billion-parameter model that it plans to release as open source, along with the dataset and training stack.

Loop's codebase is available on GitHub as a developer preview, according to the company.

Published

Corrections: AIM corrects errors of fact and notes the change on the story. Spotted one? Tell the editors. Read our editorial standards.

More in this story

AI Agents · 5 stories
  1. Infosys, Anthropic Partner to Build AI Agents for Compliance Heavy Sectors
  2. Genpact Launches Agentic AI Banking Suite for Anti-Money Laundering Investigations
  3. You are here: Soket AI Launches Loop, Open-Source Harness for Long-Running AI Agents
  4. NVIDIA Expands Nemotron AI Models for Data Centres and Desktop Devices
  5. Does the World Need a New GitHub?
About the author

Journalist

Ridhika is a journalist covering the real-world impact of artificial intelligence at the intersection of tech, policy, and society.

All stories
Belamy · Daily

The day’s AI news, before your first meeting.