We're excited to announce Agent Versioning & Sharing—a version history for every agent, the same two-key Experimental mode you already use, and unlisted share links to hand an agent to another workspace. If you've used these on skills, you already know how they work.
The Same Proven Model, Now for Agents
When we shipped versioning, releases, Experimental mode, and share links for skills, the goal was simple: bring the safety and control you expect from shipping code to the instructions your team runs in Claude Code. That model worked. So we built it on a shared internal core and extended it to agents—with commands and workflows next.
Nothing here is a new concept to learn. An agent gets the exact same mental model a skill does:
| Git | Agent versioning |
|---|---|
| Working directory | Draft — the live, always-editable agent |
| A commit | Release — a frozen, immutable snapshot |
git commit | New Release — freeze the draft into a release |
A stable tag | The stable pointer — the one release everyone runs |
git restore | Restore to draft — bring an old release back to edit |
Release History for Every Agent
An agent always has one editable draft. A release is a named, immutable snapshot of that draft—a title (required, up to 80 characters) plus optional markdown release notes. Releasing copies the draft; your draft keeps living and drifting ahead, just like a working directory after a commit. The result is an append-only history: v1…vN, each with its date and author, that you can always look back at.
Stable Is What Your Team Runs
One release per agent is marked stable. That's the version Claude Code actually loads for your team. Your draft can be a pile of in-progress edits—it doesn't matter, because the stable release is what gets served.
The first release you ever cut auto-promotes to stable, so an agent becomes servable the moment it has its first release.
Promote to Ship, Promote to Roll Back
The stable pointer is movable. Promote points it at any release:
- Point it at your newest release → that's a normal ship.
- Point it at an older release → that's an instant rollback.
Either way it only moves the pointer—instantly. Your draft and your full release history are untouched. Broke something? Promote last week's release and you're back to safe in one click.
Restore an Old Release to Keep Building
Promote changes what runs. Restore to draft changes what you edit. Restore copies a chosen release's snapshot back into your live draft so you can evolve a known-good version forward—say, take v2, restore it, and build it into v6. (Because it overwrites your current draft, we confirm before it runs.)
These are independent actions: v2 can be the stable release and be restored into your draft at the same time.
No Surprise Drafts in Production
Here's the safety guarantee that makes all of this trustworthy: an unreleased agent is never served.
If an agent has no promoted stable release, Claude Code simply skips it—it does not fall back to your draft. So a half-built agent can sit in your workspace all day without ever reaching your team. An agent becomes servable only once it has a stable release.
Dogfooding Drafts Safely: Experimental Mode
Sometimes you do want to run a draft agent against real Claude Code sessions before releasing it. That's what Experimental mode is for, and it's protected by the same two-key gate you already use for skills—both keys must be turned:
- Flag the agent experimental — a per-agent toggle.
- Turn on Experimental mode in the CLI — the single global "Experimental mode" toggle in
superclawd config, off by default. It's the same toggle that gates experimental skills—one switch for the whole machine.
Only when both are on does that agent serve its live draft—and only on that one machine. Every other machine, and every other agent, keeps running the frozen stable release. So one developer can dogfood one risky agent against real sessions while the rest of the team stays on stable.
When the draft is good: cut a New Release, Promote it, and unflag experimental.
Sharing: One Link, Another Workspace
A great agent you built shouldn't be trapped in your workspace. Agent share links let you hand one out via a single unlisted, revocable URL—the same sharing model skills already have.
A share link hands out a released snapshot, never your live draft. When you create one, you choose how it tracks your releases:
| Mode | Recipients get | Behavior |
|---|---|---|
| Pinned | a specific frozen release | Frozen to whatever was stable when you made the link. Your later releases don't change it. |
| Live | your current stable | New visitors always get your latest stable release. |
You can optionally set an expiry date and a max imports cap, and you can revoke any link at any time to kill it immediately. Mix and match: a link that expires in 7 days, caps at 10 imports, and can be pulled the moment you change your mind.
The Recipient Flow
The recipient page is shared across skills and agents—the same safe, preview-first experience:
- The recipient opens your link and sees a preview of exactly what they'd import.
- To import, they sign in and pick a target workspace.
- SuperClawd creates a brand-new, independent copy of the agent in their workspace—fully their own, with no link back to your workspace data.
Example: A Safe Iteration Loop
- Edit your agent's draft freely all week.
- Flag it experimental and turn on Experimental mode on your machine to dogfood the draft live.
- Happy with it? Cut Release v5 with notes describing what changed.
- Promote v5 to stable—now the whole team runs it.
- A bug slips through? Promote v4 to roll back instantly while you fix the draft.
- Want to share it outside your team? Create a share link, pinned or live, and send the URL.
History intact, team safe, no guesswork.
Get Started
Agent versioning and sharing are available now in all SuperClawd workspaces.
- Go to app.superclawd.com
- Open an agent and click Versions
- Cut your first Release and watch it auto-promote to stable
Want the full mechanics—the stable pointer, promote/rollback, restore-to-draft, the two-key experimental gate, and share-link modes? Read the docs at docs.superclawd.com.
Pro tip: This is the same model your skills already run, now extended to agents on a shared core—with commands and workflows next. Cut a release before each workday's edits and future-you will love having a clean point to roll back to.