Labelbox raises $100M+ Series E and pivots from data labelling to selling AI agents

What did Labelbox announce?

Labelbox said on 14 October 2026 that it had raised more than $100 million in Series E financing, including both equity and a credit facility, at a $1.25 billion pre-money valuation. The round is led by Infinitum Partners, with participation from Wipro Ventures, Andreessen Horowitz, B Capital, Tau Ventures and Snowpoint Ventures among others. Co-founder and chief executive Manu Sharma announced it in a blog post that doubles as a strategy statement.

Labelbox says it reached profitability this year while growing more than 5x year over year, and that it will use the money to put more research and engineering behind its two newer products and hire the people who deploy them. Sharma also flagged compute access as central to the next chapter, noting Infinitum’s John Yetimoglu was an early Groq investor now involved in GroqCloud.

The framing matters as much as the number. Labelbox is best known as a data-labelling company, the kind of unglamorous infrastructure that fed the first wave of machine learning. The Series E is pitched as the moment it stops being that and becomes a company that sells agents and the systems that make them improve.

From labelling data to selling agents

Sharma describes three chapters. The first was data: tools to capture expert judgment and turn it into training data, which Labelbox began selling after its 2018 founding by Sharma, Brian Rieger and Dan Rasmuson. The second was environments: as reinforcement learning from human feedback made models broadly useful, Labelbox built Horizon, RL environments that frontier labs use to train models by having them practise realistic work rather than copy examples.

The third chapter, which Sharma says just launched, is agents. Labelbox built internal agents that start on their own, work inside the tools the company already uses and improve each run, then packaged that system as Recursion Managed Agents. The through-line is the company’s founding claim that an AI system is only as good as the signal it learns from — now applied to agents rather than to models alone.

That is a crowded place to stand. Selling reinforcement-learning environments and grading harnesses is the same business several others are chasing, from Hugging Face turning math proofs into FineEnvs RL environments to Anthropic opening dynamic workflows on its Managed Agents. Labelbox’s bet is that owning the data, the environments and the agent loop together is the advantage.

What are Recursion Managed Agents?

Recursion Managed Agents, introduced on 8 October 2026, are agents a company configures with a goal, a trigger and a rubric rather than a script. A coordinator plans the work and brings in specialist agents as needed, each on whichever model — Claude, GPT, Gemini or open weights — it does best, and a separate grader agent scores every run against a plain-language checklist the team writes. When a run misses, the agent revises.

The pitch is that assistants still wait for a person at the keyboard, so work stops at night and on weekends, whereas managed agents start from schedules, alerts, Slack messages, webhooks and API calls and run all week. Labelbox says the agents work across more than 100 business apps and any MCP server, with every tool scoped as read-only, data-changing or destructive and credentials kept in a vault the model never sees. Where there is no API, an agent gets its own browser and every click is saved as a screenshot.

The learning loop is the part Labelbox treats as its real intellectual property. Each finished session leaves a short report; once five are ready, a memory run verifies what they teach and updates the agent’s memory, which the next session reads before it acts. Labelbox says it runs more than 50 such agents across its own engineering, research, sales, finance and operations.

What is Horizon, and who uses it?

Horizon is the other half of the business, aimed at the labs rather than the enterprise. Its product page describes software-generated RL environments and evaluations for reasoning, tool use and computer use, plus a simulator called WorldSim that recreates enterprise settings — GitLab, Jira, CRM, email and chat filled with realistic business data — so a model can work a hard task end to end while verifiers score each trajectory.

Labelbox says more than 90% of leading US AI labs and hyperscalers work with it, and points to case studies including data work for a Meta reasoning benchmark. The company argues the hard part of RL is the reward signal: get the tasks, verification or difficulty progression wrong and models learn shortcuts instead of real skills, which is what its software and expert network are meant to prevent.

Horizon and Recursion share a structure. In Horizon a model attempts a task, a calibrated reward says whether it succeeded, and the model improves; in Recursion an agent finishes a job, a grader scores it against the team’s bar, and every graded run feeds the memory the next one starts from. Labelbox calls both forms of a learning loop and says it builds every layer — harness, sandbox, graders, memory and models — itself.

Why investors are backing data companies as agent companies

The round reads as investors betting that the companies which labelled data for the last wave are well placed to sell agents for the next one. Labelbox’s argument is that frontier models get better and cheaper every quarter and everyone rents the same ones, so the advantage is no longer the model but what a company does with it: its workflows, judgment and a system that turns them into agents that keep improving.

It lands in a busy funding cycle for AI infrastructure and agents. The same day brought a $60 million seed for the agent startup Hone and a $55 million Series B for the chemistry-AI firm Albert Invent, while recent weeks saw raises such as Arena’s $200 million round and Alignment Index. The money is chasing the same thesis from different angles: that enterprises will run fleets of agents and need something to train, grade and govern them.

Whether Labelbox’s all-layers approach wins is unproven. Owning the data, the environments and the agent loop could compound, or it could spread the company thin against specialists in each layer. The Series E buys time and compute to find out.

What is unverified

Most of the quantitative claims here are Labelbox’s own and are not independently audited:

• The $1.25 billion pre-money valuation, the profitability claim and the “more than 5x” growth figure come from Labelbox’s announcement, not from filings.
• The round is described as “more than $100 million” including a credit facility, so the equity portion is not broken out.
• Adoption figures — 400+ customers, 90%+ of leading US AI labs, 50+ internal agents, an expert network of more than 4 million people — are company-stated.
• Recursion and Horizon capability claims, including the learning loop and app integrations, are from Labelbox’s product pages and have not been independently tested here.
• Investor roles beyond the named lead and participants, and the exact terms of the credit facility, are not disclosed.