AI 開發工具
Anthropic’s `ant apply` Brings Agents, Skills, and Memory Stores Under Declarative Version Control
The Claude Platform CLI adds `ant apply`, which synchronizes managed agent resources from Markdown, YAML, or JSON files in a repository. It provides previews, dependency ordering, and drift detection, but its state file still contains organization identifiers, and it currently manages only the Anthropic platform.
Anthropic has added `ant apply` to Claude Platform CLI 1.30.0, bringing the control plane for managed agents into an infrastructure-as-code-style workflow. Developers can declare agents, environments, skills, memory stores, and deployments in a repository, then apply either individual files or an entire directory. The CLI first generates a plan of resources to create or update and calls the API only after interactive confirmation; `--dry-run` can also be used to check differences without applying them.
Resources do not require remote IDs to be hard-coded in their configuration and can instead reference one another using relative paths. The tool resolves dependencies, creates objects in the correct order, and then records the actual IDs, versions, workspace, and local and remote hashes in `claude-lock.json`. Subsequent runs can use the two sets of hashes to detect file modifications or out-of-band changes made through the console, reducing the likelihood that individual developers and CI pipelines will create duplicate agents. Markdown frontmatter carries structured configuration, while the document body can serve directly as a system prompt or a deployment’s first message. Skills continue to use directory-based packages containing `SKILL.md`.
The technical significance is that agent prompts, tool combinations, memory, and execution environments can finally undergo pull request review, history tracking, and deployment automation together instead of being scattered across a web console. However, this is not a general-purpose Terraform provider and offers no cross-provider abstraction. Although `claude-lock.json` should not contain API keys, it still records organization, workspace, and resource IDs, so teams should assess the risk of information exposure before committing it to a public repository. Teams should also test concurrent applies by multiple users, partial state after failures, and remote drift handling, while using least-privilege credentials and mandatory dry-run reviews in CI.