- Apelogic CEO Boris Renski compares today's crowded AI agent governance market to cloud infrastructure before it consolidated
- The startup is building a governance platform based on Kubernetes and Nvidia's open source OpenShell runtime
- Governing fleets of coding agents in Jira and CI/CD pipelines has emerged as its first repeatable use case, Renski says
AI agent governance is where cloud infrastructure was in 2010, when OpenStack arrived, according to Boris Renski, CEO of startup Apelogic. Everyone agrees on the problem. Nobody agrees on the stack.
Renski should know. Apelogic is building an AI governance platform based on Nvidia OpenShell. Renski was prevously co-founder of Mirantis, an OpenStack distribution and services company that later focused on Kubernetes. He later founded FreedomFi, which sold open source gear for building 5G networks on the Helium crypto network and which became the foundation of Helium Mobile, a consumer carrier Nova Labs sold to Noble Mobile in June
I talked with Renski over lunch at a Mexican restaurant in Silicon Valley. The food was unremarkable. The conversation was outstanding.
Agentic AI trust and safety is the problem Apelogic is tackling. Enterprises, cloud providers and telcos are moving autonomous agents from experiments to production fleets that work on code, customer data and business systems. These agents need to be isolated and constrained. But the AI industry has not settled on which software should do it.
"There is a universal acceptance of the problem space," Renski said. Most vendors also agree on the basic approach: put the agent in a container or lightweight virtual machine, then define in software what it can and cannot do beyond that boundary, he said. Where companies part ways is on whose software to use to provide the guardrails, with many options.
The early cloud faced similar challenges, Renski said. Amazon Web Services (AWS) had a head start, Google and Microsoft were building their own clouds, and CloudStack, Eucalyptus, OpenNebula and OpenStack competed as open alternatives. Eventually, the market "consolidated, and some clear leaders have emerged," Renski said. He expects agent governance to follow the same path, and he is betting his company on how.
Wait, wasn't OpenStack a failure?
NASA and Rackspace launched OpenStack in 2010 to compete with AWS. That didn't work.
But OpenStack has succeeded in its own niche, providing private cloud services for telcos and government users. Broadcom's licensing changes after it acquired VMware renewed interest. The Open Infrastructure Foundation, which oversees OpenStack, forecasts that VMware migrations and workloads returning from public cloud will quadruple OpenStack deployments by 2029.
What does the OpenStack comparison predict for agent governance?
Hyperscalers are playing the role AWS played in 2010: early, well-funded and tied to their own platforms. AWS made Policy in Amazon Bedrock AgentCore generally available in March. It intercepts agent tool calls at a gateway and checks each against rules written in AWS's open source Cedar policy language. Microsoft offers Agent 365 and Entra Agent ID to discover agents and manage their identities. Google made Agent Sandbox on Google Kubernetes Engine generally available in May.
Hyperscalers' governance tools are designed for their own stacks, Renski said. For organizations outside a single cloud, including telcos, neoclouds and enterprises running private or hybrid infrastructure, Renski expects an open option to emerge. He also expects that option to follow the infrastructure agents already use rather than build a parallel universe, as OpenStack did. That means the governance layer needs to be Kubernetes-based.
That is why Apelogic builds on OpenShell, which Renski said is trying to become the agent governance software most closely aligned with Kubernetes.
What is OpenShell?
OpenShell is open source software that runs each AI agent inside an isolated container or virtual machine and enforces written policies governing which files the agent can read, which network destinations it can reach, which credentials it can use and which AI models it can call.
Nvidia introduced OpenShell at its GTC conference in March, alongside the NemoClaw reference stack for the OpenClaw agent. Because agents wrap large language models that do not behave deterministically, organizations need a way to confine them, much like putting them in jail, Renski said. OpenShell sandboxes start with minimal outbound network access.
OpenShell has assembled a support coalition. Red Hat is a contributor and maintainer, and it plans to integrate OpenShell into its Red Hat AI platform. Canonical announced an Ubuntu package in June. Nvidia said OpenShell is coming to Windows using new Microsoft security capabilities. Cisco, CrowdStrike, Google, Microsoft Security and TrendAI said in March they were building OpenShell compatibility into their security tools.
Why are coding agents the first use case?
Apelogic users are deploying the software to ensure security and auditability for agents embedded in the software development lifecycle. Mature engineering teams are moving beyond individual developers working with Claude Code or Codex, Renski said. Instead, they place agents in Jira or Linear as developer, tester and deployer agents, and in continuous integration and delivery (CI/CD) pipelines, where agents write release notes and review pull requests. Tickets are assigned to agents, which can reassign them to subagents.
"This results in a zoo of long-running agents with different roles," Renski said. Security teams want each agent held within a defined envelope that sets how many tokens it can spend, which models it can use and which tools it can reach, he wrote.
The tooling is arriving. Atlassian made agents in Jira generally available in May for Jira Cloud Standard, Premium and Enterprise, letting teams assign work items to agents, including GitHub's Copilot coding agent.
So are the risks. In June, Microsoft researchers showed that instructions hidden in GitHub pull requests could manipulate CI/CD agents into leaking secrets, Red Hat engineers noted in a July post.
What could go wrong?
The OpenStack parallel carries a warning. The biggest winners in cloud were proprietary platforms, and many companies that bet on the open alternative did not survive consolidation.
Apelogic is small, with seven employees, including Renski, and is funded by his two previous exits plus customer revenue. Red Hat plans native OpenShell support, and each hyperscaler is building governance into its cloud. Buyers are also still sorting out how to secure a growing population of non-human identities before choosing any governance layer.
For now, OpenShell is further along in ambition than in maturity. Nvidia's own GitHub page describes OpenShell as "alpha software — single-player mode," with support for the multi-tenant deployments enterprises need still ahead. Renski said the market will decide which governance stack survives, and that it has not decided yet.
Learn more about artificial intelligence
Nvidia GTC: OpenClaw is the new Linux, says Jensen Huang
Andrew Yang’s Noble Mobile acquires Helium Mobile
Zero trust for AI agents: SASE vendors race to secure non human users