Anthropic's ant apply manages Claude agents and skills as code
Terraform-style: reconcile agents, skills, memory stores and deployments from repo files, with a plan and a lockfile that blocks silent Console drift.
Copy markdown
Your agent stack, defined in files
ant apply (CLI 1.30.0+) reads agents, skills, environments, memory stores, and deployments as Markdown, YAML, or JSON in your repo, prints a create/update plan, and applies it on approval. One command reconciles the whole stack against the Claude API.
A lockfile that catches Console drift
claude-lock.json stores each resource's ID plus a local and a remote hash. Edit, archive, or delete something in the Console and the next apply refuses with 'This plan cannot be applied' until you pass --force, so your repo stays the source of truth.
Wire it into CI this weekend
Resources reference each other by relative path and apply in dependency order, and --dry-run posts the plan on a pull request. Put your agents in git, review changes in PRs, and roll back an agent by reverting a commit.
Elsewhere: open weights are 9 points off the frontier
Artificial Analysis puts GLM-5.3 and Kimi K3 at 44 on its Intelligence Index against 53 for Claude Fable 5.1 and GPT-6 Astra. Both leaders are free to download and self-host, which is most of the case for paying frontier prices at all.