Lor turns a researcher into a conductor of hundreds of AI agents

Mirendil launches Lor, a platform where hundreds of AI agents program, run experiments, and operate machines for several days.

One hundred and two AI agents, one hundred PyTorch pull requests, and an average of eight minutes of work per environment created. With Lor, Mirendil introduces a platform designed to automate a portion of the work typically performed by artificial intelligence research teams. Rather than occasionally assisting a developer, Lor allows for delegating entire experiments, distributing operations among multiple agents, and automatically spinning up the required machines.

The system is currently available in private beta, free by invitation. It runs on macOS, Linux, and the web, with the option to use models from OpenAI, Anthropic, Google, DeepSeek, and Mirendil. The company presents it as an infrastructure capable of handling much of the operational work of an AI lab, from code writing to results analysis. Hundreds of agents for a single project

An AI research experiment can require dozens of relatively independent operations: inspecting execution logs, debugging scripts, comparing configurations, reproducing results, or preparing training environments.

These tasks are often parallelizable, but coordinating them takes time. Lor allows a lead agent to distribute the work among multiple agents, each able to have its own environment and resources.

The user can thus ask to run twelve ablation experiments overnight, compare the results, and generate a report. They can also hand over the audit of a software repository, along with the creation of pull requests to fix the identified issues.

The platform supports hundreds of concurrent agents, whose activities can be monitored and paused individually.

In an internal demonstration, Mirendil states it deployed 102 agents to build environments from 100 PyTorch project pull requests, with an average duration of eight minutes per environment.

This experiment illustrates the platform's ability to distribute numerous operations. However, it does not provide an independent success rate, nor a direct comparison with a human team performing the same tasks. Agents that reserve their own GPUs

One of the main differences with traditional coding assistants relates to compute management.

In Lor, agents can request virtual machines, prepare their environment, install necessary dependencies, launch an experiment, and release the resources once the work is completed.

The user does not need to separately create an account with a cloud provider for each operation. The infrastructure is integrated into the platform, with resources that agents mobilize on demand.

This setup makes it possible to consider tasks such as reproducing a scientific benchmark table on eight GPUs, comparing several training configurations, or launching an evaluation campaign.

The agent can organize the machines, distribute the compute, and gather the results before submitting its analysis.

Mirendil has not yet specified the compute quotas associated with the private beta, nor the pricing terms that will replace free access. The provision of GPU resources therefore does not mean that users have unlimited capacity. Working for several days without keeping your computer on

Lor relies on a separation between the agent and the machine on which its commands are executed.

The agent remains active in the cloud, while its operations can run on a computer managed by Mirendil, a personal workstation, or another machine registered by the user.

This architecture makes it possible to launch a task from a laptop, have the compute executed on remote servers, and review the results later from another device.

The user can close their computer without interrupting agents working on machines that remain accessible.

Mirendil claims that certain agents can continue their activities for several days, or even weeks, by monitoring processes, identifying errors, and attempting to resolve bottlenecks.

The company notably cites an overnight session during which its agents reportedly detected and fixed vulnerabilities in LiteLLM.

This is an example communicated by Mirendil. The announcement does not provide an external audit report to verify the scope of these fixes or their definitive integration into the project concerned. A platform that can use multiple model families

Lor is not limited to models developed by Mirendil.

The platform announces support for the Claude, GPT-6, Gemini, Grok, GLM, Kimi, and DeepSeek families, as well as models from Meta.

This approach allows choosing different systems depending on the nature of the work: programming, reasoning, document analysis, investigation, or processing long histories.

Lor also integrates Fern, Mirendil's first model, capable of processing up to 10 million context tokens.

This capacity can be particularly useful for examining large software repositories, reading long agent traces, or analyzing several days of training logs without having to systematically reduce the available information.

However, a distinction must be made between context capacity and practical performance. A model accepting ten million tokens does not guarantee that it exploits every piece of information contained in that window with identical accuracy. The published evaluations for Fern show, in particular, that results depend on the difficulty of the requested reasoning. A model with hundreds of billions of parameters built by one researcher

The presentation of Lor comes a few days after the publication of another experiment conducted at Mirendil.

The lab claims that a single researcher, assisted by AI agents, developed Mirendil-Base in three and a half months, for a total compute cost of less than $10 million.

The project started from a blank repository. Agents reportedly wrote over 99% of the code, built the training tools, developed the data pipelines, ran the experiments, and helped diagnose failures.

The human researcher remained in charge of scientific direction, the choice of experiments, and the validation of results.

The model belongs to a category between 100 and 500 billion parameters, according to the range published by