The Anthropic Stack Has Five Tiers. Here's What Each One Unlocks.

Anthropic hasn't been building a chatbot.

They've been building an operating layer for work. Five products. Five tiers. Each one engineered for a different class of problem. Anthropic's own Claude Code product page and engineering write-up on Managed Agents give you the view from their side. This post gives you the view from the deployment end.

Most people are using one of them. Some are using two. The ladder exists whether you're on it or not.

The Operating Layer Most People Miss

The mistake is treating Claude as a tool you open when you need it. That's Level 1 thinking applied to a system that goes to Level 5.

Each level in the Anthropic stack is a different relationship with AI. Not better or worse. Different. Level 1 keeps you in the loop. Level 5 takes the loop out entirely. Between those two points are three rungs most businesses skip past without realising they exist.

This post maps all five. Each section covers what the level is, one concrete example of what it makes possible, and who it's designed for.

Why the Ladder Matters

Different levels produce different outcomes. This isn't marketing language. It's a structural truth about how these products are built.

Real-time help and autonomous production agents require different infrastructure. Anthropic built separate products for a reason. The same approach that works for deskside collaboration will actively limit you when you want work running without a human present.

Knowing the ladder means knowing what you're actually asking for when something on your current setup isn't working. Nine times out of ten, the answer isn't "Claude isn't good enough." It's "you're using the wrong level for the job."

Level 1, Cowork: The Deskside Agent

Cowork is Claude working alongside you, in real time, in a shared workspace your team can access.

You bring the task. Claude works with you on it. Your team shares context through the same environment. The human stays in the loop throughout.

This is the right level for reactive work, collaborative drafting, decision support, and research where the direction shifts based on what you find. It's where most teams start because it maps directly to how knowledge work already gets done.

The GA release for enterprise teams added production-grade admin controls, audit logging, and shared workspace structure. That makes this more than individual Claude usage scaled to a team. For a full breakdown of what changed at GA, see Claude Cowork Goes Enterprise: What Actually Changed.

The ceiling here is human initiation. Cowork doesn't run without you. When you're done for the day, it's done too.

Level 2, Scheduled Tasks: Set It and Leave It

Scheduled Tasks is the first rung where Claude runs without you starting it.

You set up a task, define when it runs, and step back. Claude executes on the schedule you set. You review the output when it's ready.

The use case is anything recurring that follows a consistent pattern. A weekly competitive review. A daily summary of inbound client communication. A fortnightly content brief compiled from your monitoring feeds.

These tasks don't need you to initiate them. They need you to design them well once. After that, they run.

Level 2 is where AI starts compounding. You're not spending time on the task anymore. The output shows up in your workflow, generated and ready.

The constraint is that Scheduled Tasks are still tied to your configured environment. The next two levels take that constraint away.

Level 3, Dispatch: Your Phone as a Trigger

Dispatch moves the trigger point to wherever you are.

Instead of being at your desk to initiate work, your phone becomes the remote. You send a message, a voice note, or a quick instruction from wherever you are. Dispatch takes that and kicks off work that runs in your environment without you needing to be there to execute it.

The practical value is speed and location independence. You're in a client meeting and a question comes up that needs a quick analysis pulling from your internal data. You dispatch the job mid-meeting. By the time you're back at your desk, the output is there.

This level is particularly useful for operators who move between contexts frequently. Consultants, agency owners, anyone whose working day isn't desk-bound. The phone stops being a consumption device and becomes a control interface for work that runs elsewhere.

Level 3 doesn't remove the human from the loop. It removes the requirement to be in a specific place to start the loop.

Level 4, Routines: Cloud-Native, Always On

Routines are where the infrastructure shifts.

The first three levels assume some connection to your machine or environment. Routines run in Anthropic's cloud. They're not tied to your laptop, your network, or your local setup. They run on schedules, respond to API triggers, or fire based on events in connected systems.

The practical difference: a routine can run at 3am. It can fire when a new record appears in a CRM. It can respond to a webhook from a form submission. No human trigger required.

This is where AI automation stops being a productivity tool and starts being infrastructure. Routines handle the background layer of a business operation. The data processing, the monitoring, the status updates, the routine compliance checks that currently require either a human's time or a custom-built automation stack.

Post 4 in this series covers Routines in detail: Claude Routines: Cloud-Native Automation for Business Operations.

The key word for Routines is "always on." It runs whether or not anyone is looking.

Level 5, Managed Agents: Production at Scale

This is the top of the stack.

Managed Agents is the full production agent runtime. Agents are stateful, meaning they carry context across a long-running task without losing their place. They can run for hours. They can spawn sub-agents to handle parallel workstreams and recombine the outputs. They operate in client environments, not just your own.

The class of work this makes possible is fundamentally different. An agent that researches a market, compiles a brief, generates a first-cut analysis, flags the gaps for human review, and delivers the result into your workflow is not a chatbot session extended over time. It's a different operating model.

The orchestration layer is built in. You design the agent's scope and decision logic. Anthropic's infrastructure runs it. You review outputs and handle the edge cases the agent escalates.

For more on the production release and what the API changes for teams building at this level, see Anthropic's Managed Agents API: What the Production Release Actually Changes.

This is where agents stop being experiments and become delivery infrastructure.

Where Most Businesses Are Right Now

Honestly: Level 1 or Level 2.

That's not a failure. It's the appropriate starting point. You don't learn to design production agents before you've built solid instincts for what good AI output looks like at the single-task level.

The businesses getting the most from AI right now are not the ones that jumped to Level 5. They're the ones that got Level 1 right. Shared context, consistent use, built into daily delivery. Then they asked what the next level would actually solve for them.

Some of those businesses went to Level 2 and found that was the ceiling they needed. Others had recurring workflows that clearly belonged on a schedule. The right level is the one that removes the friction costing you the most right now.

How to Think About Climbing

Don't skip levels. This is the most common mistake.

Teams that jump from Level 1 directly to production agents without the operational discipline that Levels 2, 3, and 4 build tend to get unreliable results. Not because the technology isn't capable, but because they haven't developed the instinct for agent design that comes from working at simpler levels first.

Each level teaches you something. Level 2 teaches you to design a task clearly enough that it can run without you. Level 3 teaches you to initiate work concisely. Level 4 teaches you to think about triggers and events, not just tasks. By the time you're building production agents at Level 5, those skills are what separate a working deployment from a broken one.

Start at the level that fits your current work. Use it until you've hit its ceiling. Then ask: what would the next level unlock?

Go Deeper with the Interactive Guide

The Five Levels interactive guide at /guides/claude-stack walks every level in depth. It covers the specific capabilities, the use cases, the setup requirements, and how to assess where you are in the stack. If you're not sure which level maps to your next priority, that's where to go.

If you're ready to talk about where your operation sits and what the next level of AI deployment looks like for your specific business, start with the IP Assessment. It takes ten minutes and gives you a clear read on where the opportunity is.