Agent Deployment Modes

Agent deployment modes are the practical choices for where an AI agent or automation runs and whether the live reasoning loop is still present after launch ^[src-086].

Key points

• Nate Herk frames deployment with two axes: local versus cloud runtime, and deterministic workflow versus full agentic loop ^[src-086].
• Local /loop keeps the current Claude Code session, tools, skills, context, and live agent loop, making it useful for short-lived babysitting, polling, and iteration ^[src-086].
• Claude Code scheduled tasks run from the desktop app and can persist across restarts, while cloud routines run on Anthropic infrastructure without the user's machine being awake ^[src-086].
• Cloud routines preserve the Claude Code agent pattern but require cloud-safe setup: repo context, environment variables, trigger choice, run limits, and explicit persistence ^[src-086].
• Modal and trigger.dev are deployment platforms for scripts, workflows, and tools. They are often cheaper and more deterministic, but they do not automatically include the live Claude Code session ^[src-086].
• Embedding an agent SDK into Modal or trigger.dev can restore an agent loop, but then session identity, memory, tool permissions, cost, and billing move into the developer's architecture ^[src-086].
• Hooks are the deterministic event-driven complement to deployment modes: they run before or after tool use, session start/end, or notifications without asking the model to decide ^[src-086].

Related entities

• Claude Code
• trigger.dev
• Claude Managed Agents

Related concepts

• Claude Code Loop Skill
• Claude Code Scheduled Tasks
• Claude Code Cloud Routines
• Agentic Build / Deploy Boundary
• WAT Framework (Workflows-Agent-Tools)

Source references

• ^[src-086] Nate Herk — "I Tested 3 Ways to Deploy Claude Agents (Here's When to Use Each)" (2026-05-15)