Routines and headless

Routines (Managed Infrastructure)

• Saved prompt + repositories + connectors that run on Anthropix managed infrastructure
• Triggered by: cron schedule, HTTP POST to API endpoint, or GitHub events (e. g. , new pull requests)
• No need to maintain machines or workflow files
• Created via web interface or /schedule command in Cloud Code
• Limitations: Research preview with evolving behavior; recurring schedules run at most hourly; fresh clone from default branch each run; can only push to cloud/ prefix branches by default

Headless Mode (Full Control)

• Run Cloud Code from your own scripts and applications
• Use -P flag for one-shot command execution with no interactive UI
• Reads from stdin, writes to stdout—pipes like standard shell tools
• Skips auto-discovery of hooks, skills, plugins, MCP servers, and cloud MD file
• Faster startup; only loads explicitly allowed tools
• Useful when job needs your environment or custom logic

Structured Output

• Pair JSON schema with output format JSON to constrain output structure
• Schema-matched object lands in structured output field of JSON response
• Extract with jq command and pipe to database or another script

Multi-Step Automation

• Capture session ID from JSON output to resume later
• One script kicks off work; another resumes with full context

Agent SDK

• Library that embeds Cloud Code in TypeScript or Python applications
• Exposes query function and same primitives as schema
• Pass prompt + options (allow tools, system prompt, permission mode)
• Iterate over streamed messages from Cloud

When to Use Each

• Routines: Repeat work on managed infrastructure
• Headless (-P): Pipe data through scripts
• Headless (--bear): CI needs consistent results per run
• Agent SDK: Embed work inside your own product

Takeaways

• Routines run on managed infrastructure triggered by schedules, webhooks, or GitHub events—no machine maintenance needed, but limited to hourly recurrence and cloud/ prefix branches
• Headless mode (-P flag) lets you pipe Cloud Code through scripts like standard shell tools, skipping auto-discovery for faster startup and custom environment control
• Structured output uses JSON schemas to constrain responses into a structured output field that you can extract with jq and pipe to databases or other tools
• Multi-step automation captures session IDs to resume work later, enabling one script to kick off a job and another to continue with full context
• Agent SDK embeds Cloud Code in TypeScript or Python applications via a query function, giving you programmatic control over prompts, tools, and permissions