[PRIVATE PREVIEW]

A distributed filesystem for trillions of files.

Fast, highly scalable shared storage that lets trillions of agents store and share files across sessions concurrently.

Tens of thousands of agents writing to one shared drive

[SHARED FILESYSTEM FOR AGENTIC WORKLOADS]

A shared memory layer for AI agents

Agent Drive behaves like a shared cloud filesystem mounted directly into your sandbox's file tree. Hand work between agents, share context and artefacts, and let multi-agents collaborate without a sync layer.

Compare storage options
  • Many agents, one drive

    Concurrent read-write access from multiple sandboxes at once, so a fleet of agents can work on the exact same files.

  • Mount as a path

    A POSIX-compliant FUSE client makes a drive appear as a normal directory inside every sandbox that mounts it.

  • Bottomless storage

    Drives grow automatically with no fixed capacity limit. Nothing to pre-provision.

[BUILT FOR TRILLIONS OF AGENTS]

A filesystem, not an object store

Object storage such as S3 was not designed for POSIX operations, or for the many-small-files pattern of agentics. Agent Drive is built as a filesystem from the ground up, designed to hold trillions of files and keep every read fast.

  • O(1) reads

    Reading a file costs a single disk seek, however many files the drive holds.

  • Small-file optimized

    Designed for the patterns of agentic workloads: many small reads and writes at very high concurrency.

  • Custom FUSE client

    An optimized, POSIX-compliant FUSE client runs in every sandbox, developed in-house for agent workloads.

  • Scale-out by design

    Metadata and file content scale independently and horizontally, so a drive grows without re-sharding.

  • Built-in replication

    Data is replicated across nodes, so a drive stays durable and readable through node failures.

  • Faster than S3

    Agent Drive is 5–10x faster on small-file I/O than object-storage-based filesystems.

“Blaxel Agent Drive is 5-10x faster than the S3-based filesystem solution we had previously implemented.”
Dirk BreeuwerCo-founder, Corvera

[FINE-GRAINED DATA ACCESS CONTROL]

Share the drive, but not every file

Label-based ACLs, evaluated outside of the sandbox, decide which workloads can reach a drive, which folder they can access, and whether they can write or read-only.

Read the permissions docs
  • Label-based rules

    Match on workload identity labels such as team=backend or env=prod.

  • Read or read-write

    Set a mode per rule, so one agent writes while another only reads.

  • Path scoping

    Confine a workload to one folder, such as /reports, or even one file.

  • Update rules live

    Change permissions on a drive that is already in use.

[S3-COMPATIBLE ENDPOINT]

Reach your drives from outside Blaxel

Every drive exposes an S3-compatible HTTP endpoint. Read and write its files from any S3 client, such as the AWS CLI, boto3, or the AWS SDK for JavaScript, without mounting the drive in a sandbox first.

Read the S3 endpoint docs
  • Standard S3 clients

    Point any S3-compatible tool at the drive's endpoint, with path-style addressing.

  • SigV4 authentication

    Requests are SigV4-signed with a service account's API key, using path-style addressing.

  • Seed and collect data

    Load datasets before agents start, and pull their results into your own pipelines.

[PURPOSE-BUILT FOR AGENTS]

Build a shared memory layer for your agents

  • Handoff of small files

    Agents produce many small files: notes, JSON artifacts, tool outputs. Agent Drive is tuned for this access pattern, with high concurrency and no serialization step in between.

  • Shared memory

    Keep long-term agent context, conversation history, and accumulated knowledge across sessions. New sandboxes mount the drive and pick up where previous ones left off.

  • Dependencies

    Put code dependencies or datasets on a drive. Every sandbox that mounts it gets immediate access without downloading them again.

  • Parallel agents

    Fan out work to dozens of sandboxes running in parallel, each writing results to the same drive. A coordinator agent reads the aggregated results when the workers finish.

Get early access to Agent Drive.

Apply for the private preview to use a distributed filesystem for your AI agents. We're onboarding new teams every week.