Labelbox has announced more than $100 million in Series E financing, combining equity with a credit facility, at a $1.25 billion pre-money valuation. Its announcement dated October 14 names Infinitum Partners as lead investor, alongside Wipro Ventures, a16z, B Capital, Tau Ventures and Snowpoint Ventures.
The company plans to put the financing into research, engineering, customer-facing teams and computing capacity for Horizon and Recursion. The package is not described as entirely equity funding, and the announcement does not disclose the split between equity and credit. It also does not set out a separate India expansion plan.
From training environments to workplace agents
The two products address different parts of AI development. Horizon provides training and evaluation environments for models that must reason, use software tools or operate a computer. Labelbox describes its WorldSim component as a way to recreate business systems such as GitLab, Jira, customer-management software, email and chat using realistic business data. Scenarios can change the task conditions, including simulated outages or changes to a workflow.
That approach makes the testing environment part of the product. A model has to complete work across tools rather than produce a plausible answer to a single question. Labelbox says the environments combine checks of the result with human comparisons of alternative attempts. The stated aim is to reward useful performance without teaching shortcuts. These are the company's descriptions of its training infrastructure, not an independent finding that a model will perform reliably in every workplace.
Recursion Managed Agents, introduced on October 8, is the enterprise execution side of that business. Teams specify a goal, a trigger and criteria for an acceptable result. A coordinator can divide the assignment among specialist agents, while a separate grader checks the evidence against the team's criteria. Labelbox says sessions can begin from a schedule, an alert, a message or an API call.
The launch description also sets out limits on that autonomy. Tools have distinct permissions, and the company says work that repeatedly fails is stopped for review rather than presented as complete. A session record includes the model interactions, tool calls, produced files, cost and grading. For a team reviewing an automated assignment, that record offers more to inspect than the final document alone. The launch's demonstration scenarios should not be read as independently verified customer outcomes.
Evaluation and access controls remain part of the work
The underlying Recursion platform predates this month's financing. In its June 24 introduction, Labelbox described a system connecting task environments, evaluation and reinforcement learning. Workflows, policies and unusual cases become executable training situations. Evaluation considers intermediate decisions and execution quality as well as the final result, producing feedback that can be used in later model training.
This helps explain what the new capital is intended to support beyond running more chatbot sessions. The earlier platform description treats the record of an attempted task as training material: graded sequences can feed fine-tuning or reinforcement learning for specialist models. It is a proposed route from workplace expertise to more tailored systems. Whether a particular deployment improves still needs testing against that organisation's own tasks; the description is not a guarantee of accuracy or lower costs.
Security is another part of that deployment question. In an April 10 engineering discussion, Labelbox described using temporary access rather than broadly persistent credentials, and granting permissions for a specific task and period. It also said it had introduced a waiting period before adopting new software dependencies, giving security teams time to identify risks in newly released packages.
Those controls concern how software and people reach sensitive systems, separate from how an AI result is graded. The company described simulation exercises to identify which credentials could be exposed in a hypothetical compromised-package incident. Its account provides background on its approach to access and software supply chains. It does not establish that autonomous agents are free of security risks, nor does the funding announcement supply an independent audit of the newly described product capabilities.
Related coverage: Dalimss News has reported on RapidCanvas's enterprise AI funding and RevRag AI's in-app agents.
