Zeroset Raises $5.2M to Build an AI Agent Context Layer

A five-person startup that launched in January is betting that the missing piece of enterprise AI is not a smarter model but a better map of the company itself. Zeroset, founded by two college dropouts, emerged from stealth this month with $5.2 million in pre-seed funding to build an AI agent context layer that learns how individual businesses actually operate.

The round was co-led by Gradient Ventures and 2048 Ventures, with Leblon Capital participating, according to the company's announcement. Zeroset's first product, called Nebula, is now in a closed research preview and aims to capture the sequence of tasks, decisions, and activities flowing through an organization's workplace software, giving AI agents a persistent model of the environment they are supposed to automate.

The funding round

Zeroset was co-founded by Akshat Kannan, who left Stanford University this year, and William Zhang, who left the University of Texas in 2025, according to reporting by Pulse2. The San Francisco startup keeps a deliberately small team and plans to spend the new capital on hiring researchers and engineers, supporting early business deployments, and covering the heavy cost of model training.

Training those models is expensive enough that the budget line matters on its own. One training run costs roughly $100,000 over two weeks, according to coverage of the announcement. That detail hints at the ambition behind Nebula: the company is not fine-tuning prompts or bolting retrieval onto chatbots. It is training its own models on the traces of work itself.

No valuation was disclosed for the round. The investors are a telling mix. Gradient Ventures backs AI-first infrastructure from inside Google's orbit, while 2048 Ventures focuses on early-stage technical founders. Their joint presence signals that both the AI-infrastructure crowd and the early-stage seed community now see the enterprise context problem as a standalone category worth funding.

Why enterprise agents keep stalling

Ask a modern AI agent to draft a recap of a meeting and it will produce something competent in seconds. Ask it to carry a product request from approval to release, figuring out whether last week's sign-off still applies to this week's changes, and the same agent is likely to stall a few steps in. That gap is the problem Zeroset's AI agent context layer is designed to close.

The thesis, as the company describes it, is straightforward: off-the-shelf models absorb public knowledge from the internet but know nothing about the ins and outs of any specific company. Nobody's documentation explains that a claims analyst at one insurer checks an internal dashboard, reads the related email thread, and then updates the CRM record in that exact order, with exceptions that live only in habit. Until that knowledge is captured somewhere an agent can use it, enterprise agents keep failing on the multi-step workflows that matter.

Nebula approaches the problem by connecting to workplace software and modeling the flow of work over time, turning observed patterns into repeatable playbooks. According to Founderland, the system links tools including Microsoft 365, SharePoint, Outlook, Teams, and GitHub. Rather than retrieving an answer when an agent asks a question, the AI agent context layer creates what Zeroset calls a state layer, helping agents behave consistently over weeks of work instead of answering each query in isolation.

That distinction between retrieving information and understanding a process is the core product bet. Document search can tell an agent what the expense policy says. It cannot tell the agent how expense reports were actually approved in this company last quarter, which manager was the bottleneck, or what exceptions were tolerated. An AI agent context layer built from the traces of real work is meant to answer those second-order questions automatically.

A crowded race for agent memory

Zeroset is not alone in chasing the agent-memory problem. Mem0 and Zep are the most frequently cited competitors, and both offer memory and retrieval infrastructure for AI applications. What distinguishes Nebula, if the company's framing holds, is that it models enterprise dynamics rather than storing facts for lookup. The company describes the result as enterprise worlds: learned representations of a company's state and how that state changes in response to actions.

The company compares the challenge to robotics, as reported by Pulse2. Increasingly capable robots did not become useful from better policies and more compute alone; they needed representations of their surroundings and an understanding of how those environments responded to action. Zeroset wants to apply the same lesson to businesses, giving agents a working model of the office the way a robot has a working model of the room.

Commercially, the bet is that an AI agent context layer becomes the connective tissue for enterprise automation, priced as a combination of license fees and usage charges tied to data volume and agent activity, according to the AI Market Watch report. That pricing structure aligns the cost of the layer with the volume of successful automation it enables, which is the right shape if agents actually get more reliable.

Whether that reliability materializes is still an open question. Nebula's closed research preview means its ability to reduce human involvement remains unproven in real deployments, as early coverage has noted. Early enterprise customers will need to test whether agents using the AI agent context layer follow approval steps correctly and require fewer interventions than agents working from retrieval alone.

For the broader agent ecosystem, though, the round is another signal that the frontier of agent infrastructure is moving below the model layer. Earlier this week, Hedera launched agent accounts to give agents non-custodial spending ability, and DigitalOcean rolled out agent learning services to turn agents into self-improvers. Zeroset's AI agent context layer fits the same pattern: agents are getting the tooling to be workers rather than demos.

As reported by Founderland, the company declined to share a valuation and is hiring as it expands access to Nebula's research preview. Gradient Ventures published its own announcement of the investment, describing the startup's work on modeling long-horizon enterprise workflows, and Pulse2 covered the pre-seed round on October 7. If the AI agent context layer thesis proves out, the companies that teach agents how offices really run may matter as much as the labs that build the models.