Open source veteran Brian Douglas has officially launched Paper Compute Company, a new startup aimed at addressing a critical gap in the burgeoning field of AI agents: the underlying infrastructure required for their robust deployment and management in production environments. Douglas, known for his significant contributions to open source communities including GitHub and the Cloud Native Computing Foundation (CNCF), articulated this vision on LinkedIn, stating, "Everyone’s building agents, not enough people are building the systems underneath them." This observation forms the bedrock of Paper Compute’s mission, which is to provide the essential, often overlooked, infrastructure that allows AI agents to operate reliably and securely at scale.
The current landscape of AI agent development is characterized by rapid innovation in agent capabilities, but often lacks a cohesive and standardized infrastructure. Companies are increasingly deploying agents for various tasks, yet the supporting systems are frequently a patchwork of Software Development Kits (SDKs), cloud-native plumbing, and custom internal tooling. While the individual components may exist, they have not yet coalesced into a mature, integrated system capable of meeting the demands of production-grade AI applications. Paper Compute aims to fill this void by building this missing layer, with a strong commitment to an open-source foundation.
Addressing the Production Gap: Observability and Control
Douglas explained to The New Stack that the fundamental challenge for many organizations is gaining the confidence to run AI agents in diverse and demanding situations. "Having proper cloud-native tooling to control agents and LLMs in production – that’s where we’re seeing a lot of energy happening right now in the market. We weren’t seeing this three months ago," he stated, highlighting the recent acceleration in market demand for such solutions.
Paper Compute’s approach is rooted in a series of open-source product releases that lay the groundwork for their infrastructure offering. The company formally unveiled itself in March, but the foundational work began earlier. In February, they open-sourced Tapes, an observability layer designed to capture and record agent activity without requiring any modifications to the existing application code. This "zero-instrumentation" approach is crucial for seamless integration into existing workflows. Tapes operates by sitting between AI agents and their inference providers, capturing telemetry data. Its components include a proxy service to intercept and record traffic, an API server for querying session data, messages, and metadata, a command-line interface (CLI) for managing recordings and performing searches, and a terminal user interface (UI) for detailed step-by-step tracing of agent activity.
Following Tapes, Paper Compute released StereOS, a hardened Linux-based operating system specifically engineered to run AI agents within isolated, sandboxed environments. This provides development teams with enhanced control over agent behavior in production, mitigating risks associated with autonomous systems. The dual aim of these products is to tackle the core problem: once AI agents are deployed, organizations often struggle with both visibility into their operations and control over their autonomy. Douglas summarized this succinctly in the company’s March launch blog post: "Tapes show you what happened – StereOS makes sure it can’t go further than it should."
To illustrate the practical application of their technology, Paper Compute developed a Gmail agent demo. This demonstration showcased an OpenClaw agent running within a StereOS virtual machine, tasked with triaging an inbox under strict operational controls. The agent was capable of reading and classifying emails but was explicitly prevented from deleting or sending replies, and its access to external services was limited to a small, pre-defined allowlist. Every interaction of the agent was meticulously captured by Tapes, creating a complete, replayable log of prompts, responses, and decisions. This data was stored locally in a tamper-evident format. Upon completion of the session, the environment was dismantled, credentials were destroyed, and the full execution history remained accessible for audit and analysis.
A Foundation in Open Source and Enterprise Engineering
Brian Douglas brings a wealth of experience from the open-source and cloud-native ecosystems. Prior to co-founding Paper Compute, he served as GitHub’s director of developer advocacy and led ecosystem and developer experience initiatives at the CNCF. His entrepreneurial background includes founding Open Sauced, a developer insights platform acquired by the Linux Foundation, which aimed to assist contributors in discovering and engaging with open-source projects.
John McBride, Paper Compute’s CTO and co-founder, shares a similar pedigree. He previously collaborated with Douglas at the Linux Foundation and Open Sauced, where he was the lead AI engineer. McBride also has extensive experience as an engineer at Amazon Web Services (AWS), contributing to critical infrastructure projects. The growing Paper Compute team was further bolstered by the addition of founding engineer Matthew Yeazel, who spent several years at AWS working on services like EC2, Amazon Linux, and Bottlerocket, an operating system focused on container workloads.
This collective experience in open source, cloud infrastructure, and large-scale systems engineering informs Paper Compute’s strategic positioning. The company aims to occupy a crucial intersection between developer tooling and core infrastructure, providing solutions shaped by a deep understanding of operating systems and applications at scale. "We know what production infrastructure looks like – agents deserve the same rigor," Douglas emphasized.
The Emerging Data Play: From Compute to Insights
As organizations increasingly deploy AI agents into production, understanding the intricate workings of these systems becomes paramount. The sheer volume of data generated by agents is often fragmented, difficult to access, or tightly coupled to the specific frameworks used for their development, such as LangChain and its associated observability tool, LangSmith. These tools, while valuable, have historically operated within their own ecosystems, creating potential silos for data analysis.
Douglas highlighted a practical scenario: when an agent encounters an error, the immediate need is not to painstakingly reconstruct the sequence of events, but to have a readily available, comprehensive record that can be queried directly. "So instead of wondering how you got somewhere, you can just interact with the data," he explained. "There’s no SDK, no extra code – it just runs in the background."
This focus on capturing agent activity points to Paper Compute’s long-term vision: the true value lies not just in the infrastructure itself, but in the rich data it generates. Agent interactions, Douglas argues, create a detailed record of system behavior and decision-making processes that becomes a valuable asset in its own right. "The value is actually more in the data that goes through the agents," he stated.
While Paper Compute is still in its nascent stages, the company is actively developing a commercial strategy built upon its open-source offerings. At its core, Paper Compute is not merely building tools; it is establishing a foundational layer for the end-to-end management of AI agents in production, encompassing observability, execution, and eventually, orchestration.
The company’s name, Paper Compute, encapsulates this philosophy. Douglas explained that the model centers around the concept of "paying for compute," where usage and the flow of data through these systems become the primary drivers for monetization. This approach leverages the telemetry captured by tools like Tapes, transforming agent sessions into reusable "skills" and utilizing this data to optimize model deployment and execution.
Early iterations of the platform are currently undergoing internal testing, and the Paper Compute team is beginning to engage with prospective users. The company’s long-term objective is to serve teams that require stringent control over AI system operations, particularly those in regulated industries or operating in on-premise environments where security, compliance, and auditability are non-negotiable.
Douglas draws a parallel to the evolution of cloud-native infrastructure, referencing the advent of Kubernetes over a decade ago. "We had Kubernetes, like, 13 years ago, when cloud native infrastructure started to take shape," he remarked. "Now we’re seeing a similar wave with AI tooling."
The broader ecosystem is already reflecting this trend. Kubernetes is increasingly being utilized to orchestrate GPU-intensive AI workloads, extending its proven capabilities for scaling CPU-based applications to a new class of compute-bound systems. Concurrently, open standards like OpenTelemetry are evolving to capture traces, metrics, and logs from LLM and agent systems, providing a more consistent and vendor-neutral method for understanding production behavior.
Douglas anticipates that open-source development will drive the commoditization of AI infrastructure tooling, leading to increased adoption by enterprises. "Through open source, we’re going to see that tooling commoditized," he predicted. "And in the next six months, I think we’ll see enterprises – banks, large companies – ready to scale this." However, he concluded with a crucial caveat: "But the tooling still needs to give them the trust to run agents in production." Paper Compute aims to be the catalyst that provides that essential trust.