Neul Labs: Dependable AI and Agent Systems

Dipankar Sarkar
Dipankar Sarkar · · 3 min read

What Neul Labs does

I founded Neul Labs in 2024 to build and deliver the systems around an AI model: identity, authority, state, evaluation, execution boundaries, recovery, and operational ownership.

Current delivery includes agentic AI services using Azure AI Foundry and AKS/Kubernetes, custom sandboxing, and MCP in .NET. Customer names, architectures, usage scale, and security outcomes remain confidential. Earlier delivery work, before the current cloud stack, included Python and AWS-based document extraction, validation, routing, and LLM workflows for clients whose architectures and scale also remain confidential.

Alongside the open-source projects, Neul Labs also runs CloseGate, a governance product demonstrating action tiers, segregation of duties, mandatory approval for high-impact actions, materiality-based routing, and hash-chained audit trails for tool-using agents. Like the open-source projects, CloseGate is public design and delivery work; it has no published external certification or third-party deployment record.

Public reference work

The Neul Labs GitHub organisation provides inspectable implementations and research artefacts, including:

  • Regulus for runtime identity, purpose, policy, privacy, human oversight, kill switches, and evidence export.
  • fast-langgraph for durable agent state and checkpoint operations.
  • grite for Git-native multi-agent coordination events.
  • agentvfs for isolated agent workspaces and execution boundaries.
  • brat as a crash-safe multi-agent harness.

These projects establish public design and implementation work. Unless a separate primary source says otherwise, they do not establish third-party adoption, independent audit, certification, or universal performance.

One measured result comes from the Before the Pull Request study behind grite and fast-langgraph: in its measured setup, duplicate or conflicting rework among coding agents fell from 78% to 0%, useful throughput more than tripled, and replicas converged without silently dropping a valid write. That is an operation-level result from one measured setup, not a claim about every coordination workload.

The operating thesis

An agent becomes a production system when it can cause side effects. Dependability then comes from explicit permissions, durable state, idempotent actions, evaluation, observability, recovery, rollback, and an accountable human owner—not from a more persuasive demonstration.

FAQ

What does Neul Labs build? Governed AI agent infrastructure: runtime identity and policy, durable agent state and checkpointing, Git-native multi-agent coordination, isolated execution boundaries, and crash-safe multi-agent harnesses — public through the Neul Labs GitHub organisation.

Is Neul Labs’ delivery work public? Public reference implementations are open source; current customer delivery on Azure AI Foundry and AKS/Kubernetes remains confidential, including customer names, architectures, and usage scale.

What evidence backs Neul Labs’ coordination work? The Before the Pull Request study, measured against grite and fast-langgraph, where duplicate or conflicting rework fell from 78% to 0% and useful throughput more than tripled in that specific measured setup.

How does CloseGate relate to the open-source projects? CloseGate is Neul Labs’ governance product rather than an open-source library: it applies the same operating thesis — explicit permissions, durable state, and an accountable human owner — as a deployable action-tiering and approval layer, using hash-chained audit records to make each decision reviewable after the fact.

Does public code on GitHub mean Neul Labs’ systems are independently audited? No. Public reference implementations establish inspectable design and implementation work; unless a separate primary source says otherwise, they do not establish third-party adoption, independent audit, or certification.