Claude Automation Primitives — Routines, Managed Agents, Dispatch
Source: wiki synthesis: Routines, Managed Agents, Dispatch, Claude Cowork, Hermes Agent Overview, Hermes Productivity, Marketing Automation Use Cases, Claude Agent Hierarchy
Anthropic shipped three distinct async-automation primitives in a three-month window (Jan–Apr 2026). They look similar from 10,000 feet (“Claude runs tasks without me babysitting”) but differ on the dimensions that matter: where the work runs, who triggers it, what permissions it has, and which audience it targets. Picking the wrong primitive is an expensive mistake — not in dollars but in architectural rework.
The three primitives at a glance
| Primitive | Triggered by | Runs on | Trigger modes | Audience | Surface |
|---|---|---|---|---|---|
| Routines | Schedule / API call / GitHub webhook | Anthropic cloud | cron (hourly/nightly/weekly), POST to routine endpoint, GitHub events | Developers | Claude Code on the web |
| Managed Agents | Programmatic (API) | Anthropic cloud | Composable API calls; multi-agent coordination (preview) | Developers (infra-savvy) | Claude Platform API |
| Dispatch | Mobile message | User’s own desktop | Tap on phone → desktop executes | Business users | Cowork on desktop + mobile |
Orthogonal dimensions
Where the work runs
- Routines and Managed Agents run on Anthropic infrastructure. Your laptop can be closed. Scales independently of your local machine.
- Dispatch runs on your desktop. The machine must stay awake with Claude open. Computer use occupies your screen while Dispatch executes.
Who triggers it
- Routines have three trigger modes — cron schedule, API endpoint with bearer token, GitHub webhook (Claude opens one session per PR and feeds updates from that PR to the session).
- Managed Agents are triggered programmatically — you build an orchestration layer on top, they’re the runtime primitive.
- Dispatch is triggered from your phone. The “kick off from bed, come back to output” interaction pattern.
State and identity
- Routines: actions happen AS you (commits carry your GitHub user, Slack messages use your account). Each run is stateless — no state carries between runs. Uses Connectors (Slack, Linear, Jira, Drive, GitHub) rather than local MCP servers.
- Managed Agents: scoped permissions, checkpointing, execution tracing, secure sandboxing. Long-running sessions can operate autonomously for hours.
- Dispatch: operates your actual desktop. Task memory is local per machine; no cloud sync. Access governed by per-folder permission dialogs.
Pricing shape
- Routines: usage caps per plan — Pro 5/day, Max 15/day, Team/Enterprise 25/day. Additional purchased via extra usage.
- Managed Agents: consumption-based — standard Claude Platform token rates + **58/month for a 24/7 always-on agent before tokens.
- Dispatch: included in Cowork Pro (20) or Max (200).
Decision framework — pick by the question you’re answering
| If the question is… | Reach for | Why |
|---|---|---|
| ”Run this every night at 2am” | Routines — Schedule | Purpose-built for cron-shaped automation |
| ”Trigger on every PR, one session per PR” | Routines — GitHub webhook | Webhook mode specifically designed for this |
| ”Webhook from my internal alert system” | Routines — API | Each routine has an endpoint + auth token |
| ”Multi-agent coordination with checkpointing” | Managed Agents | Preview feature; no one else has it |
| ”I want to build my own agent runtime on top of Claude” | Managed Agents | That’s exactly what it is |
| ”Tell Claude from my phone to rename those files on my Mac” | Dispatch | Mobile → desktop is Dispatch’s whole point |
| ”App has no API — computer use needed” | Dispatch | Routines and Managed Agents can’t see your desktop |
| ”Continuous self-improving agent loop” | Hermes | Hermes is a different architecture — still compose with Claude primitives |
Hermes Agent context
Hermes is its own agent architecture, not a Claude primitive — but it intersects with all three:
- Hermes can invoke Managed Agents as a sub-runtime for hard long-horizon tasks (per productivity-workflows).
- Hermes can dispatch via Cowork for desktop-bound tasks that need local file/app access.
- Hermes’ own scheduled tasks can be replaced by Routines when the task fits Routine’s trigger modes and doesn’t need Hermes’ learning loop.
Anti-patterns (seen in prior research)
- Using Routines for desktop-app automation. Routines can’t operate your desktop; they only use Connectors. Use Dispatch.
- Using Dispatch for unattended overnight jobs. Desktop must stay awake; unreliable as a cron replacement. Use Routines.
- Building your own cron scaffolding around Managed Agents. Managed Agents is the runtime, not the scheduler. Wrap it with Routines (API mode) rather than bespoke orchestration.
- Using Managed Agents for a simple “run this every Monday” task. Overkill. Routines is the right altitude.
Key Takeaways
- Three primitives, three audiences. Routines → developers who want cron-shaped automation. Managed Agents → developers who need a runtime to build on. Dispatch → business users who want async delegation from mobile.
- Cloud vs desktop is the architectural fork. Routines and Managed Agents are Anthropic-hosted. Dispatch needs your machine awake. If your job touches local files or native apps, you need Dispatch; otherwise prefer the hosted primitives.
- Pick by trigger shape, not by “which one is newest.” Cron → Routines. Webhook → Routines. Programmatic from your backend → Managed Agents. Human-in-the-loop from mobile → Dispatch.
- Compose rather than replace. Hermes can call any of the three. The Advisor Strategy works inside all three. evaluator-optimizer patterns apply across all three.
- Pricing shapes decide defaults too. Routines’ per-plan caps force careful use. Managed Agents’ $0.08/session-hour makes “always-on” pricey. Dispatch is free-with-Cowork but chews your desktop.
Related
- Claude Code Routines
- Claude Managed Agents
- Dispatch in Claude Cowork
- Claude Cowork (Product Overview)
- Claude Agent Hierarchy — When to Use Which
- Agent Workflow Patterns
- Hermes Agent — distinct architecture that composes with these primitives
- AI Marketing Automation Use Cases — real workflow destinations
- Cost & Intelligence Levers — the tuning levers orthogonal to primitive choice
- All connection articles
Try It
- Inventory your existing automation. List every “runs without me” task in your stack. For each, classify against the decision table — does it match its current primitive, or would another fit better?
- Test the webhook fit for Routines. Take one GitHub workflow where you currently use CI to call out to an LLM. Port to Routines with GitHub webhook trigger. Measure whether the “one session per PR” model is materially better than stateless CI runs.
- Prototype one Managed Agent. Take a long-running task you don’t have (but have wanted) — multi-hour research, continuous monitoring, etc. Stand it up on Managed Agents. $0.08/hour adds up; measure whether the intelligence gained justifies it.
- Do one Dispatch handoff. Send your phone a task that requires your desktop’s native apps (PowerPoint export, Illustrator manipulation, whatever). Feel where the seam is between “direct connection is fast” and “computer use is slow.”
- Map your Hermes workflows against the three primitives. Anywhere Hermes runs its own scheduler, consider whether Routines would be a simpler substrate.