Company Spotlight: Zeroset Builds State for AI Agents

Zeroset is a San Francisco applied AI research company building Nebula, a state layer for long-running enterprise agents. Co-founders Akshat Kannan, CEO, and William Zhang are focused on a specific weakness in agent infrastructure: models can reason over a task while still misunderstanding the changing business state around it.

Nebula is designed to organize workflow history, decisions, ownership, handoffs, exceptions, and provenance into a queryable model of how work actually moves. The product remains in closed research preview, but Zeroset's direction matters now because enterprises are pushing agents beyond single prompts and into processes where stale context can compound into operational risk.

The company has raised $5.2M in pre-seed funding from Gradient, 2048 Ventures, and Leblon Capital. Zeroset is using that capital for model training, early enterprise deployments, and hiring across research and systems engineering.

About Zeroset

Zeroset describes itself as a small applied AI research lab. Its thesis starts with an uncomfortable fact about enterprise software: most systems record activity, but few preserve a reliable explanation of why the work changed.

A request can begin in a ticket, pick up an approval in Slack, change owners in a project tool, and end with an exception that lives only in one employee's memory. An agent may retrieve every relevant document and still act on the wrong version of reality. The missing object is not another chunk of text. It is current state.

Zeroset's first product, Nebula, is meant to capture that state. The system traces actions taken by people and agents across workflows, resolves entities and time, tracks contradictions, and returns compact context for the next action. Zeroset calls the broader idea an enterprise world model: a durable representation of how a particular organization works, not a generic model of how work should happen.

Why Nebula Matters Right Now

The enterprise AI market has become very good at selling capability. Models can draft, classify, search, code, and call tools. Production reliability is a different contest. A model can be excellent at choosing an action and still fail because it does not know which policy is current, which exception was approved, or which handoff remains unresolved.

Nebula sits between model capability and workflow execution. Zeroset argues that retrieval should be a view into memory rather than the entire memory system. Its "Memory Is Not Retrieval" memo frames memory as persistent state with explicit updates, temporal validity, provenance, and contradiction resolution.

That framing moves the product beyond a smarter search box. If the architecture works, an agent does not merely find a past approval. It can understand who approved it, which policy version applied, whether the conditions changed, and how the precedent should influence the next case.

The Product and Its Evidence

Zeroset says Nebula gives agents the current state and workflow history required to complete more work with fewer repeated mistakes and less context. The company website describes dedicated infrastructure, configurable retention, single-tenant deployments, customer-managed keys, and audit logs delivered to customer security systems.

The company's security page says infrastructure is primarily hosted on AWS in the United States, data is encrypted at rest and in transit, and row-level security separates workspaces. Zeroset also says customer data is not used to train models. A SOC 2 Type II audit is in progress, which is an important qualifier for buyers evaluating a young infrastructure vendor.

Zeroset has also published Atlas v0, a benchmark for world modeling, temporal reasoning, procedural knowledge, preference learning, and other memory capabilities. The company reports that Nebula scored 76.9% overall against a 66% average for specialized comparison systems. Those results clarify Zeroset's technical thesis, but they are company-produced evidence, not independent proof of production outcomes.

Leadership, Capital, and Focus

Kannan and Zhang are building an unusually research-heavy company at an early stage. The founders have chosen a problem where the demo is the easy part and the edge cases are the product. Enterprise state includes messy permissions, contradictory records, forgotten approvals, policy changes, and years of operational habits that no benchmark can fully simulate.

Zeroset announced a $5.2M pre-seed round in October 2026. Gradient and 2048 Ventures co-led, with Leblon Capital participating. The capital is intended for model training, early deployments, and hiring researchers and systems engineers while Nebula remains in a controlled preview.

That measured rollout is sensible. Infrastructure that shapes an agent's view of business reality cannot afford to be casually wrong. The product has to earn trust through better workflow completion, lower intervention rates, reduced repeated errors, predictable latency, and economics that improve as usage grows.

Why Zeroset's Hiring Mix Is a Market Signal

The Zeroset careers page lists full-time San Francisco roles in research, distributed systems, Rust systems engineering, and product engineering. It is a compact hiring plan, but the composition says more than a headcount target.

Research roles point toward model and memory behavior. Distributed-systems and Rust roles point toward the reliability, throughput, and control requirements of persistent infrastructure. Product engineering points toward the integration work required to turn a technical thesis into something an enterprise team can deploy and measure.

This is not evidence that Zeroset has solved the market. It is evidence that the company understands where the hard work sits. The next wave of agent infrastructure will be judged less by theatrical demos and more by whether the system survives long-running workflows without losing the plot.

What Zeroset Signals for Enterprise AI

Zeroset is competing in a crowded layer that includes memory providers, context platforms, orchestration systems, observability vendors, and infrastructure for durable agents. The category boundaries remain fluid because the underlying problem is still being decomposed in real time.

Zeroset's sharper bet is that state becomes a distinct control layer. Models may change, and tools may multiply, but the enterprise still needs a canonical view of what happened, what changed, who owns the next action, and why the workflow moved the way it did.

That makes Zeroset worth watching even before Nebula reaches general availability. The company is asking whether operating knowledge can compound as software instead of evaporating across systems and employees. If Nebula can answer that question inside real deployments, enterprise agents may finally inherit something more valuable than another prompt: an accurate sense of where the business stands.