Picture an AI agent handed a job on Monday morning — a research sweep, a security audit, a codebase migration — and still working on it on Friday, remembering everything it tried, every tool it used, every decision it made along the way. That is the kind of work most AI tools today simply are not built to hold together. Bengaluru-based Soket AI has built the layer meant to hold it: LOOP, an open-source harness that lets AI agents run for hours, days and weeks instead of minutes.
A laptop running through the night — an illustration of LOOP, Soket AI’s harness for agents that keep working long after the chat window closes.
The harness, not just the model
Soket AI, a research lab building frontier AI models and systems that institutions can operate under their own control, unveiled LOOP on 7 October 2026 at Cypher 2026 in Bengaluru. The company says it was selected under the IndiaAI Mission to build frontier foundation models for India and the Global South, and LOOP is its entry into the same class of tooling as OpenAI’s Codex and Anthropic’s Claude Code — with one decisive difference: it is open source, under the Apache 2.0 licence, and the company says developers can inspect every line of code, contribute to it, and run it anywhere.
The idea, the company says, is simple and timely. Through 2025 and 2026, AI moved from models that answer, to agents that act, to harnesses that can stay on a task — and the longer an agent runs, the more the harness matters. LOOP is available now as a developer preview on GitHub for Linux, macOS and Windows, and installs with a single command.
Built to run for weeks
LOOP is built in Rust, the company says, designed to operate close to the system layer for low latency, efficient resource use and parallel tool execution. Developers can plug in any AI model simply by adding its base URL, and LOOP handles the rest — no lock-in to a single provider’s hosted agent.
Its session management works the way developers already think. The company says sessions can be branched, paused and resumed, much like working with Git, and session history is preserved during context compaction so nothing is silently lost. A single goal can be split across a swarm of agents that share memory, with isolation options ranging from rootless containers to micro virtual machines. Agents run in secure sandboxes by default, the company says, with no access to files unless allowed. Skills and templates can be turned into simple slash commands, and LOOP supports the Model Context Protocol, so it can call tools from compatible AI clients or serve its own.
More Read
Light enough to run anywhere
The engineering shows up in the numbers the company shared at launch. In a local comparison using eight workers, with LOOP and Claude Code pointed at the same Qwen model through an OpenRouter base URL and key, the company says LOOP used 7.7 times less RAM and 17 times less local CPU. Its memory usage rose from 29 MiB to 223 MiB as workers were added, following a near-linear pattern — the kind of efficiency, the company says, that could let thousands of agents run on a small system. The measure covers the harness process itself, and the company is clear it is a resource comparison, not a quality score.
A demonstration at Cypher showed what that efficiency is for: LOOP researching the day’s conference schedule through multiple tools, then returning the talks and sessions it judged relevant — a small job, but exactly the kind of multi-step, multi-tool work that long-running agents exist to do.
For work that cannot be handed to a hosted agent
LOOP is designed, the company says, for environments where trust is non-negotiable: cybersecurity, banking, finance, legal and defence. Those are the workloads where an organisation cannot hand its data and tools to a hosted agent — where sovereign deployments need air-gapped environments, zero-trust execution and auditability, and where a task that cannot fail needs a harness built for determinism.
“AI agents are moving from answering questions to doing real work, but the infrastructure around them hasn’t caught up,” founder and chief executive Abhishek Upperwal said at the launch. “We built LOOP to give agents the ability to stay on a task, learn from what they’ve done, and keep moving forward.” His invitation to developers was characteristically open: “A strong model cannot reach peak performance by itself. 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.”
Part of a bigger frontier bet
LOOP sits inside a larger programme at Soket AI. The company says its broader focus is Math Code Reasoning agents, built on close to 30 trillion training tokens, primarily focused on mathematics and code, along with nearly two billion reasoning traces — combining mathematical reasoning, natural-language reasoning and code execution so agents can interact with their environments over longer tasks. Upperwal also announced at Cypher that the company is training a 26-billion-parameter model it plans to release as open source, along with the dataset and training stack.
More is on the roadmap for workloads that cannot fail: sandboxing personas, multi-agent systems, layered memory, remote agents, trusted remote sandboxes and group chat for agents, plus telemetry and a desktop app. Those are not part of this preview — but the preview is already out in the open, free to fork, and ready for developers to run on their own machines.
From a Bengaluru research lab, Soket AI is building the unglamorous layer underneath the AI boom — the harness that lets an agent keep going after everyone else has logged off. The models get the headlines; LOOP is betting that the harness is what turns them into real work.
Keep reading: One Prompt, One Whole Film: Inside Eloelo’s Dolphin AI, the Studio Turning India Into a Story Exporter · The Voice That Switches Languages Mid-Sentence: Inside Arrowhead AI’s Fletch · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
