Labelbox Raises Over $100M Series E at $1.25B Valuation to Advance RL and AI Agents – Unite.AI

Labelbox said on Oct. 14, 2026, that it raised more than $100 million in Series E funding at a pre-money valuation of $1.25 billion, in a round led by Infinitum Partners that combines equity capital with a line of credit.

Wipro Ventures, Andreessen Horowitz (a16z), B Capital, Tau Ventures and Snowpoint Ventures also participated, among others. CEO and co-founder Manu Sharma detailed the terms in a company blog post, and the company said the financing will support continued research and engineering on its two core products, Horizon and Recursion Managed Agents, as well as the expansion of the teams that bring those products to customers. Sharma wrote that access to computing will be central to the company’s next chapter, noting that Infinitum’s John Yetimoglu was an early investor in Groq and is now involved in building large-scale computing infrastructure at GroqCloud.

Company location and history

Labelbox said it reached profitability in 2026, growing more than fivefold year-over-year. The company has more than 400 customers, including more than 90% of major U.S. AI labs, and says its base includes frontier AI labs, hyperscalers, startups and Fortune 500 companies.

Sharma, Brian and Dan founded Labelbox in 2018. In a founder’s note, Sharma wrote that he came to the United States to learn to design and build planes and rockets, working on computational fluid dynamics and neural networks, and that his subsequent roles at DroneDeploy and Planet Labs showed him how many of machine learning’s opportunities lay in data.

The post describes the company’s history in three chapters: training data software for people who develop artificial intelligence; reinforcement learning environments built after what Sharma calls the ChatGPT moment, when more capable models began to improve by practicing real work; and agents, which the company built for itself before opening the system to customers.

Sharma also wrote that Labelbox builds every layer of its platform itself: an agent infrastructure designed for ultra-long-horizon agents, a sandbox infrastructure that manages millions of virtual machines, its own post-training infrastructure for reinforcement learning, and an expert network of more than 4 million people.

Horizon-reinforcement learning environments

According to the company, Horizon’s WorldSim system builds simulated businesses, with code, dashboards, spreadsheets, decks and documents full of realistic business data, where a model attempts a difficult multi-step task from start to finish while each step is recorded as a trajectory. Deterministic verifiers evaluate each trajectory equally on every run, and when two trajectories are both correct, an expert’s preference teaches the model what Sharma calls flavor. The company describes operation at the scale of thousands of worlds and millions of graded trajectories, tuned to a lab’s pass@k goals through coding, computing, research, multimodal, and voice work.

The Horizon product page lists domains that include autonomous AI research, agent coding and software engineering, multimodal knowledge work, using voice with agentic tools, computer use, and cybersecurity. Sharma wrote that Labelbox now works with 90% of leading frontier AI labs and hyperscalers.

Recursion managed agents

The funding announcement states that Labelbox opened Recursion Managed Agents to enterprise customers in early October 2026. By product design, a Managed Agent receives a goal, a trigger such as a schedule, an alert, a Slack message, a webhook or an API call, and a rubric that defines what good work looks like. A separate grader scores each rubric analysis line by line and indicates the evidence for each call; when work is missing, the agent reviews it.

Agents can operate in three modes: multi-agent, in which a coordinator chains together specialists who each bring their own model, instructions, tools and permissions; team, where copies of a split agent work on a shared board and check each other’s results; and auto, the default, where the agent decides on its own when a job needs specialists or a team.

The company says agents work across more than 100 enterprise applications and any MCP server, with each tool scoped read-only, data modification, or destructive, and credentials kept in a vault so the model never sees them. Sessions run in isolated sandboxes, and Labelbox says the platform works without data retention at each LLM provider. Its listed compliance program covers SOC 2 Type II, GDPR, CCPA, NIST 800-171, and PCI DSS. Pricing is set per agent session hour, with compute and model spend combined into a single estimate, and existing Labelbox enterprise customers receive a one-time credit of $100.

Describing the company’s internal launch in a product launch post dated October 8, 2026, Labelbox wrote: “The surprise wasn’t that agents could do the job. It was how quickly they improved.” The company reports managing 51 agents across 11 teams, spanning software engineering, reliability, security, machine learning research, evaluations, product, finance, sales, strategy and procurement.

In an internal example, Labelbox said its pull-request review agent ran 39 graded reviews on a single version and model of the agent, and that, among the 33 reviews that passed the grader, the average cost dropped from $13.15 for the first 11 reviews to $8.90 for the last 11, 32% less per review, while the agent’s memory was updated 15 times and the time per review was approximately 27 minutes. According to the product’s memory mechanisms, once five new session reports exist, and at most once every hour, a curator agent reads them and updates the agent’s memory; customers can browse what an agent has learned and eliminate any lessons that no longer apply.