`ant apply` ships — Claude agents and skills deployable from YAML
Claude Platform CLI v1.30.0 adds ant apply, which syncs agents, environments, skills, memory stores, and deployments declared as YAML files — and generates a claude-lock.json lockfile to track resource versions across deploys. This is the same infrastructure-as-code pattern that made Terraform the default for cloud infrastructure: you write what you want, the tool diffs and applies. If you're managing Policy Assistant's protocol definitions through the dashboard or by hand, this is the upgrade path to reproducible, version-tracked agent configs. Worth an afternoon to convert one agent definition to YAML and see if the schema maps cleanly onto your protocols structure.
▸Do it — 4 steps in Policy Assistant agent definitions60 min
- Update the
ant CLI to ≥ 1.30.0 (check the install docs at platform.claude.com) - Run
ant apply --help to see which resource types (agents, skills, memory-stores, environments, deployments) map to YAML - Export one existing agent definition and rewrite it as a
.yaml file - Run
ant apply and inspect claude-lock.json — confirm version tracking works before migrating the rest
Fable 5.1's published system prompt: copyright walls up, harm reduction URLs in
Simon Willison dissected Anthropic's updated published system prompt for Fable 5.1. Two changes that matter for anyone building on the API: copyright limits are now broader — Claude refuses to reproduce even portions of songs, poems, or books through any method, including code-generated SVGs and ASCII art — and when asked to generate something resembling protected content, it creates an original design rather than refusing flat. Test any content pipelines you have against this. The tone guideline 'accountability without self-abasement' is now explicit text in the prompt: Claude acknowledges errors without over-apologizing, which is the behavior you want in an agent that raises policy conflicts or surfaces evaluation failures. The notable gap: substantial sections are still unpublished, including the full end_conversation tool spec — the published prompt is significantly smaller than the live one.
Anthropic's Model Hardware Standard: typed read/write for agents touching real devices
Anthropic's research preview defines a specification for how AI agents should discover, communicate with, and control physical laboratory equipment — using read/write primitives, natural-language device documentation, and MCP as the transport. Partners include Genentech and Carnegie Mellon. The part that generalizes beyond biotech: the sandboxing model. Agents get typed, declared access — no side channels — and the boundary is the interface contract, not the model. This is the reference implementation of 'agents with physical-world access done safely.' If your interest in agent safety and sandboxing is as an engineering problem, this is what a production answer looks like. And since it's built on MCP, the same tool-calling pattern Policy Assistant already uses would be the integration surface.
↗Read — anthropic.com6 min
- Agents interface with devices via
read and write primitives — no side channels; the boundary is the typed interface contract - Device discovery works in standard formats: each device publishes natural-language documentation of its capabilities and parameters
- Multi-device orchestration is built-in: agents can sequence steps across instruments and autonomously recover from errors
- Built on MCP — same tool-calling pattern Claude Code already uses
- Research preview only: early partners are biotech and quantum computing labs, not general software toolchains yet
Why read anyway — The security model section shows exactly how Anthropic is thinking about 'agents with physical access' — the typed-interface approach translates directly to any agent with real-world tool access, not just lab robots. original ↗