Snapshot any bucket
Branch it instantly
Recover anything
Every agent and experiment gets its own fork of real, current data. The source is never touched.
- Zero-copy forks, with no storage until they diverge
- Immutable snapshots and point-in-time recovery
- Promote through review, or discard at zero cost
Copy it, fake it, or risk itNone of them work
Manual dataset copies
Copies are slow, expensive, and stale before the run starts.
Synthetic data
Safe, and wrong where it matters. It skips the edge cases your model hits first.
Runs on production
Overwrite production and the old state is gone, with no record of what ran.
Branch replaces all three.
Snapshot, branch, experimentPromote or discard
The source is never touched, and every action is logged automatically.
Snapshot
Capture a bucket’s exact state at a point in time. Instant, zero-copy, and read-only.
Branch
Fork a writable branch from any snapshot or from live state. Copy-on-write, so storage is zero until it diverges.
Experiment
Run pipelines, migrations, or training on the branch. Many branches run in parallel and the source is untouched.
Promote or discard
Promote through a governed workflow, or discard at zero cost. Every action is logged automatically.
What this makes possible
Agent sandboxing
An autonomous agent gets its own branch to read, write, and delete freely. A bad run is discarded at zero cost.
Replace staging
A branch of production is current and costs nothing until it diverges. No refresh pipeline, nothing to drift.
Shadow testing and evals
Run a new model or agent on real data, side by side with production. Promote only the version that wins.
Safe bulk changes
Migrations, cleanup, deduplication, redaction. Run them on a branch, review the diff, then promote.
Reproducible training
Train on today’s production state, then pin the snapshot so the run reproduces later.
Point-in-time recovery
Restore a bucket to the moment before a bad write or delete, whether a person or an agent made it.
Nothing ships unreviewedWho signs off is up to you
IronShard shows exactly what changed: added, modified, removed, plus cost and policy impact. Promote automatically when policy checks pass, or require a human sign-off.
Agents experiment on branches. Production lives in managed buckets.