What happened
Mirendil announced Lor on October 28, 2026, describing it as a workspace intended to support the work of an AI lab. The company says Lor agents can write code, run experiments, and iterate on results over days, using any chosen machine and operating at scales from dozens to hundreds of agents. Lor is now in private beta: access is by invitation through a waitlist, and the service is free during that beta.
The product’s central claim is that a single job can be split among many agents. Mirendil describes a workflow in which one agent distributes work to dozens of others, while users can follow each agent’s progress and stop individual agents. Its example is a run that built environments from 100 PyTorch pull requests using 102 agents, at an average of eight minutes per agent. The announcement frames parallel work as useful for reading agent traces, conducting experiments, and building training environments from real pull requests.
Lor also includes managed compute. According to Mirendil, agents can request nodes, set them up, run work, and release them afterward, without the user supplying a cloud account or doing initial infrastructure setup. The company says agents receive full control of the virtual machines and nodes they bring up. Proposed uses include experiments, training runs, and evaluation sweeps. This makes compute allocation part of the agent workflow rather than a separate handoff to an operator.
The workspace is designed for asynchronous work that continues after a user closes a laptop. Agents run in the cloud, while commands can be sent to a machine selected by the user: a Lor-managed cloud computer, a laptop, or another always-on machine. Mirendil says agents can work only on machines the user has added and that machines can be removed at any time. It gives examples such as overnight training-run supervision, converting feedback into fixes, and a claimed overnight LiteLLM security-bug finding and repair.
Lor supports Mirendil’s Fern alongside a list of third-party and open models, including Anthropic Claude, OpenAI GPT-6, Google Gemini, xAI Grok, Meta Muse, Z.ai GLM, Moonshot AI Kimi, and DeepSeek models. Fern is described as supporting up to 10 million tokens of context. Lor is available on macOS, Linux, and the web; Mirendil says Windows, iOS, and Android are planned later. The announcement also says a single Mirendil researcher used agents to pretrain Mirendil-Base, with the project taking three and a half months, but it does not provide independently comparable details about that work.
What the archive adds
This is a new access and product event on a known model story. The earlier Field Note, Fern, recorded Mirendil’s claim of a 10M-token context window. Lor is now presented as the workspace where that model can be used, although access remains invite-only.
Trend signal
Lor supports the pattern that harness architecture outlives model choice. Its pitch is not one model alone: it combines a model roster with routing work across agents, machines, and provisioned compute. The announcement is evidence of product direction, not proof that this integrated harness is reliable or preferable in production.
Connections to other events
The Univer Workspace Field Note described a shared surface where people and agents create, collaborate, and review work. Lor addresses a different surface: long-running technical execution across machines and compute. Both treat the workspace as more than a chat window, but Lor’s announcement emphasizes experiments, code, and infrastructure operations.
Why it matters
Teams building agentic systems often assemble several layers themselves: a model gateway, a queue or orchestrator, machine access, cloud credentials, observability, and a mechanism for keeping work alive after a developer disconnects. Lor proposes to collapse much of that assembly into one product. If it works as described, the practical advantage is not merely parallel prompts; it is assigning bounded work that may need GPU capacity, retries, remote commands, and human interruption controls.
The trade-off is correspondingly serious. A workspace that can provision nodes and dispatch commands to selected machines becomes part of a team’s operational control plane. Builders evaluating it should distinguish the appeal of a 10M-token model context from the separate questions of access control, cost after beta, auditability, reliability of self-unblocking behavior, and portability of workflows outside the service.
Limits and uncertainty
The announcement is Mirendil’s product description, and it does not establish independent performance, reliability, security properties, or total cost of Lor. No post-beta pricing, usage limits, model-specific pricing, service-level commitments, or detailed permission model is provided. The examples of agent runs and security fixes are company examples rather than reproducible benchmarks. “Any machine” is qualified by the requirement that the machine be added by the user, but the source does not explain credential handling, isolation, logging, review gates, or what actions agents may take without confirmation. Fern access through Lor is stated, yet the terms and availability of that access are not detailed.
Skill to build now
Map one long-running task in your own workflow before adopting a workspace: define its finish condition, the machines it genuinely needs, the maximum parallelism, the checkpoints requiring review, and the authority an agent may hold. That exercise separates a useful asynchronous job from a vague request to “keep working,” and it makes a future Lor evaluation more concrete.
Reader takeaway
Lor is Mirendil’s first stated route to Fern and a private-beta attempt to package multi-agent coordination, machine dispatch, and compute provisioning into one workspace. Its promise is operational leverage for long-running technical work; its unanswered questions are the controls, economics, and evidence required before that promise can be treated as production infrastructure.
Open question
Can Lor demonstrate that agents provisioning compute and operating across user-selected machines remain inspectable, permissioned, and economically predictable once the free private beta ends?