Mirendil has opened a private beta of Lor, a platform for AI agents to run coding, research and model-building tasks for days at a time. Its notable promise is to coordinate many agents and provision computing resources for them, putting ambitious work within reach of smaller teams, at least in principle.
Mirendil Watch analysis
What happened
Lor is available by invitation and is free during the private beta. Mirendil says it works on macOS, Linux and the web, and can run several frontier and open-source models. The company also describes Fern, its own model with a context window of up to 10 million tokens.
The platform is designed for work that does not end with a quick chat: agents can run experiments, write code and keep working asynchronously. Users can divide a job across dozens or even hundreds of agents, follow their progress and stop them. Lor can also bring up computing resources on demand and release them when work is finished. Agents may work on a Lor-managed computer or another machine the user has added.
Mirendil says one researcher used agents to help pretrain its coding model, Mirendil-Base, in three and a half months. That is the company’s account of its own work, not an independent assessment of how readily other researchers can reproduce it. The company’s Lor announcement gives examples including running overnight experiments, auditing a code repository and reproducing a paper’s results.
Why it matters
Lor is aimed at a more demanding use of AI agents than drafting a paragraph or answering a question. If agents can handle long-running work, divide it into parallel tasks and call on GPUs when needed, they could help researchers and engineers attempt projects that would otherwise require more people or substantial infrastructure.
The practical test is whether people can steer and check that work without becoming full-time supervisors of an ever-expanding agent workforce. Mirendil says users can follow agents’ progress and stop them, and can direct them to machines they have added. Those are meaningful control features, but the announcement’s examples do not establish how reliably the system completes difficult jobs.
Our read
This is a substantial AI product launch, not just another chatbot wearing a lab coat. Lor’s ambition is to make long-running, parallel AI work easier to organise, with compute provision built into the platform. That makes it worth watching, especially for small research teams and developers who cannot casually assemble a large cluster.
For now, treat the striking examples as Mirendil’s demonstrations of what it wants Lor to do, rather than a performance guarantee. The useful question is whether users can inspect results, catch mistakes and get dependable work without the agents quietly turning a complicated task into a complicated bill.
What to watch
- How access to the invite-only beta expands, and what pricing follows it.
- Whether independent users can reproduce the company’s long-running research and coding examples.
- How much control users have over compute use, costs and agent activity.
- Whether support for other platforms, including Windows, iOS and Android, arrives as planned.
Discussion spark: Would you trust a team of AI agents with a research or coding project lasting days if you could monitor and stop them, or should long-running work require closer human supervision?
Sources and evidence
- Source update (28 October 2026, 00:00 UTC)
not affiliated with or endorsed by Mirendil