Paper Compute: The Infrastructure Layer AI Agents Are Missing

Every major AI lab is shipping agents. Fewer teams are building the infrastructure that makes those agents auditable, bounded, and safe to run in production. Paper Compute is building that layer.

The observation behind Paper Compute is precise: agent development has outpaced agent infrastructure. Teams are deploying multi-step autonomous systems that call external APIs, read sensitive data, and take actions on behalf of users — with minimal tooling for understanding what those agents did or preventing them from doing more than they should. Co-founder Brian Douglas, former GitHub Director of Developer Advocacy and contributor to CNCF, frames the gap directly: “Everyone’s building agents, not enough people are building the systems underneath them.”

The company’s first open-source release, Tapes (February 2026), addresses the observability side. Tapes operates as a proxy service with an API server, CLI, and terminal UI — inserting itself between the agent and whatever inference API it calls. No instrumentation changes are required in the agent code itself. Every session produces a durable, structured record of what the agent requested, what it received, and how it responded. The design is deliberately local-first and works with major inference providers including OpenAI and Anthropic.

The second product, StereOS, takes a different approach — not observation, but constraint. StereOS is a hardened Linux operating system designed for sandboxed agent execution with strict environmental controls. Where Tapes provides the audit trail, StereOS defines the execution boundary. In a Gmail agent demo, the system could read and classify emails but was structurally prevented from deleting messages or accessing services not explicitly listed in its permission scope. The Paper Compute framing captures the layered design: “Tapes show you what happened — StereOS makes sure it can’t go further than it should.”

CTO John McBride and co-founder Matthew Yeazel bring infrastructure depth from AWS EC2, Amazon Linux, and Bottlerocket. That lineage matters: the problems Paper Compute is solving are closer to OS and systems engineering than to ML engineering, and the team’s background reflects that distinction. The company is developing a pay-per-compute pricing model, suggesting infrastructure-as-a-service positioning — billing follows compute consumption, not seat count.

The gap Paper Compute is targeting has been visible since AI agent observability became a named problem in the cloud-native community. Jaeger v2 addresses distributed tracing for agent execution paths; Tapes addresses the inference-layer record. They solve adjacent parts of the same production readiness problem. For engineering teams moving agents from demo to deployment, the infrastructure question is no longer hypothetical — and cloud-native AI production patterns have yet to converge on a standard stack for agent execution.

Two products, two primitives: know what your agent did, and prevent it from doing more than it should. That’s a minimal but non-trivial baseline for production agent deployments, and it’s the baseline most teams currently lack.

Photo by Brett Sayles on Pexels

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.