Articles
Guides and deep dives on storage for AI agents and parallel AI workloads: copy-on-write branching, conflict-safe publishing, sandbox persistence, and filesystem architecture.
What is agent-native storage?
Agent-native storage gives AI agents a real filesystem with the guarantees fleets need: instant isolated forks, versioned history, conflict-safe publishing, and provenance. Here's how it works and why object stores and shared drives fall short.
Copy-on-write branching for AI training data
Copy-on-write branching lets you fork terabyte-scale datasets in milliseconds without duplicating a byte — isolated, writable, and mergeable. How CoW works, why AI training and evals need it, and what it replaces.
A shared filesystem for AI agents — without handing out S3 keys
Putting cloud credentials inside agent processes is how buckets get deleted. How a brokered mount model gives every agent full filesystem access with short-lived, workspace-scoped grants — and no S3 keys in the sandbox.
Conflict-safe publishing for agent fleets
When parallel agents write to shared storage, last-writer-wins silently destroys work. How compare-and-swap publishing, isolated sessions, and per-commit provenance let hundreds of agents work on the same data without losing a byte.
s3fs alternatives: mounting S3 as a filesystem in 2026
s3fs makes a bucket look like a folder — until small files, renames, and concurrent writers expose the object store underneath. An honest tour of the alternatives: goofys, Mountpoint for S3, JuiceFS, EFS, Archil, and Amulet.
Persistent storage for AI sandboxes (E2B, Daytona, Modal)
Sandboxes are disposable by design — which makes agent state a problem. Pause/resume, volumes, and S3 sync all have sharp edges. How a mounted, versioned workspace gives every sandbox durable state that outlives it.