Keep Claude Code for the work. Deploy the skill as a team workflow.

Claude Code gives a person good visibility into their own work. /workflows shows each phase with agent prompts, results and tool calls; cloud scheduled tasks keep a run history with full transcripts; Team and Enterprise cloud sessions can be shared with the team; dashboards report usage, adoption and contributions; and OpenTelemetry exports events, with traces in beta.^{[1, 6, 10, 17, 18]} What each run leaves is a trajectory, read one session at a time. The reviewed docs do not describe one history, across the organization, of typed step outputs per run.^{[1, 17]}

An Avalanche workflow produces structured runs. Each step is a checkpoint with typed inputs and outputs, and Avalanche Cloud keeps every run of every deployed workflow in one run history, by workflow and version.^{[41, 42]} Because the steps are the same on every run, you can open step three on one run, or compare step three across a thousand runs: which step fails most, which outputs drift, what a new version changed. The analogy is the shift data teams made from running one-off queries to managed transformations in Snowflake or Databricks.

Governance follows from that: one catalog of workflows, one version history and one run history across every teammate's agent, instead of skills copied between people and sessions. Avalanche Cloud adds its own controls: SSO, organization roles and per-action permissions, scoped, encrypted secrets, sandboxed steps and an audit trail of who promoted or rolled back each deployment. Claude Code still leads on full audit logs, its Compliance API and spend controls; Avalanche Cloud has no spend limits yet.^{[19, 20, 21, 41]}

Sources ^{[1, 6, 10, 17, 18, 41, 42]}

Yes. Teams commit skills to .claude/skills/ in a repository, and a git repository with a marketplace file distributes plugins to the whole team. Team and Enterprise organizations can also sync plugins from private repositories under Organization settings, Plugins & skills.^{[2, 3, 5]} Managed settings can require or block plugins, MCP servers, hooks and permissions, and the Team and Enterprise plans list SSO, spend controls and, on Enterprise, roles, SCIM, audit logs and a Compliance API.^{[19, 20, 21]}

What those features share is a package that each person's install loads. A plugin update reaches a user only when its version string changes, and a plugin loaded from a local path reads the current files at every session start.^[4] Anthropic's docs describe no deployed workflow artifact with its own id, no rule that keeps a running workflow on the version it started with, and no single run history across the team for one workflow version.^{[1, 4]} That gap is the job Avalanche is built for.

Sources ^{[2, 3, 4, 5, 19, 20, 21]}

Cloud scheduled tasks in Claude Code normally clone the repository's current default branch for each run, and a run marked Succeeded means the session ended without a reported failure, not that the work was done.^[6] That is reasonable for a person checking a transcript. It is harder when ten people's agents rely on the same process.

In Avalanche Cloud, a deploy freezes the workflow from committed source into a checksum-verified version with an id. Runs already in flight keep the code and bundled dependencies they started on; new runs use the new version. If a release goes wrong, instant rollback makes any previous version of the whole workflow live again with no rebuild, and promoting a version forward works the same way.^[41] Avalanche Cloud is in early access through a waitlist.^[47]

Sources ^{[6, 41, 47]}

Claude Code dynamic workflows save completed agent results and reuse them when you relaunch the run from the same session. The docs are precise about the limit: if a script starts agents A, B, C and D and B fails, the relaunch returns A from cache and runs B, C and D again, even if C and D had finished. A fresh session starts over, and a workflow takes no human input mid-run; for sign-off between stages, Anthropic suggests a separate workflow per stage.^[1] Separately, /rewind restores tracked file edits and conversation in a session.^[7]

In Avalanche every step is a checkpoint. On Avalanche Cloud, a retry starts at the failed step and keeps the accepted outputs and committed workspace files of every step before it, so a document run that takes hours or days does not repeat work that already passed its checks. Where a result needs sign-off, the run waits for an owner's approval inside the same workflow instead of a separate workflow per stage.^{[41, 42]} Agent steps declare typed outputs, and inline results are validated against that type before they reach the next step.^[40]

Sources ^{[1, 7, 40, 42]}

Yes, and that is the intended setup. Today you install the Avalanche skill in Claude Code with npx skills add Trampoline-AI/avalanche, describe the outcome you want, and Claude writes the workflow as Python your team reviews in git.^[43] Once deployed, the workflow is something Claude and every teammate's agent can call.

An agent step can also run Claude through the Claude Agent SDK, keeping its signature, typed outputs and tools, with the step itself checkpointed (pull request #124), and Agent Skills written as SKILL.md folders load inside those steps (pull request #136).^{[45, 46]} This is our own integration work, not an Anthropic partnership.

Sources ^{[43, 45, 46]}