Writing

Claude Cowork vs Claude Code: The Automations I Actually Run

The difference between Claude Cowork and Claude Code, where a Claude automation actually runs, and the three I use every week. From someone who started on Zapier.

August 18, 2026 · 10 min read · Sebastian Pinzon Duran

I started with Zapier. Then Make. Then n8n, which is where most of my heavier work still lives.

Before any of that I was writing Apps Script inside Google Sheets to build client reports, because the alternative was rebuilding the same deck every Monday for a year.

None of that was glamorous. All of it worked, right up until it did not.

What has changed in the last two years is not that automation became possible. It is that building one stopped requiring the person who knows the tool. That is the part I did not see coming, and it is why I now watch people who have never opened a builder in their life ship something useful in an afternoon.

What actually changed

For twenty years, automating something meant listing every path it could take, then building each one by hand.

That is why so many automations die quietly. Not because the logic was wrong. Because reality showed up in a shape nobody drew.

Three generations of automation Scripts run one fixed path until they hit a hard stop. Workflow builders require every branch to be drawn by hand. Agents choose their route at runtime. THREE GENERATIONS OF AUTOMATION 01 Scripts One fixed path 02 Builders Every branch by hand 03 Agents Chooses at runtime
Twenty years of automation in three moves. Only the last one decides anything.

Zapier and Make and n8n were a real gain. They made the branching visible, and a visible branch is a fixable branch.

But the rule was the same across all three: if you did not draw the branch, it did not exist.

An agent works differently. You give it the goal and the tools, and it decides the steps at runtime. When something arrives in a shape you did not anticipate, it does not stop. It reads it, makes a call, and tells you what it did.

That is better. It is also the new risk. A script that breaks costs you a morning. An agent that decides wrong costs you whatever it touched.

The old tools did not die. They became hands.

This is the part I did not see coming either.

n8n, Make and Apify all ship MCP servers now, which means Claude or ChatGPT can call them directly. n8n's is first party, in public preview, built into every edition including the free self-hosted one. Make turns your scenarios into tools an AI can trigger with structured inputs. Apify's hosted server lets an agent search their store for a scraper and run it.

So I did not throw anything away. I stopped opening the canvas. The scenario I built two years ago still does the work, and something else decides when to call it.

Which gives you a simple rule for where to put each step.

Use the deterministic tool when the outcome should never vary. Move this row to that sheet. Same result every time, and when it breaks you can see exactly which step broke.

Use an agent when the step needs a decision. It can handle the case you did not plan for. It can also be wrong about it.

Most of what I run is both. Deterministic pipes moving the data, an agent in the middle doing the one part that needs reading.

Claude gives you two doors into this

Both run on the same engine. They are built for different people.

Claude Cowork is the agentic workspace inside the Claude app. No terminal, no code. You connect folders and it works on your real files, opens your browser through Claude in Chrome, and hands back documents: spreadsheets with working formulas, decks, organized folders. Desktop, web and mobile on paid plans.

This is the one for knowledge workers, which is most people reading this. Finance, legal, ops, research, marketing. Worth saying that the line is moving, though. I came up through agencies on the marketing side and I now spend my week in a terminal, and I am not the only one. Plenty of marketers are turning technical because the tools stopped requiring permission from engineering.

Claude Code is the same engine pointed at the machine itself. Repos, servers, databases, scripts, terminal. It is what I use to build and deploy things.

The distinction people get wrong is the next one.

Where the automation runs: your machine or Anthropic's cloud Two products sit at the top: Cowork, which has no terminal and works on your files, and Claude Code, which is a terminal and works on your repos. Below them a heavy divider splits the canvas into two zones. In the your machine zone: a Routine set to Local, minimum one minute, machine must be on; and slash loop, which dies with the session. In the Anthropic's cloud zone: a Routine set to Cloud, minimum one hour, laptop can be closed; and Cowork's slash schedule, where the computer can sleep. A Routine is one concept in Claude Code and you pick Local or Cloud when you create it. Cowork's scheduled tasks run in the cloud only, so where a job runs is a separate question from which product you use. COWORK No terminal. Your files. CLAUDE CODE Terminal. Your repos. YOUR MACHINE Routine LOCAL 1 MINUTE MACHINE MUST BE ON /loop DIES WITH THE SESSION ANTHROPIC'S CLOUD Routine CLOUD 1 HOUR LAPTOP CAN BE CLOSED Cowork /schedule COMPUTER CAN SLEEP Where it runs is the real choice.
The product you use and the place the automation runs are two separate questions.

Where it runs matters more than which one you pick

I assumed Cowork was the cloud one and Claude Code was the local one. Neither is true.

In Cowork you type /schedule inside a task, and the scheduled version runs in the cloud, so your computer does not need to be awake for it. That is the case even though Cowork itself is sitting on your desktop working through your local folders.

In Claude Code you create a routine and pick where it runs. Same idea, two locations.

Cloud. It clones your repo, uses your connectors, and fires on a schedule, an HTTP call, or a GitHub event. Minimum interval is one hour. Your laptop can be closed. Still in research preview, so expect the edges to move.

Local. The same routine running on your own machine, which means it reaches your files and your tools, and it goes down to a one minute interval. The machine has to be on.

There is a third thing, /loop, and it is not a routine. It runs inside an open session and dies when the session ends. Use it to watch a job finish, not to run your week.

So the choice is not which product. It is what the job needs to touch. If it needs a file on your laptop, it runs on your laptop. If it needs to happen at 7am whether or not you opened the machine, it runs in the cloud.

Three automations I actually run

None of these are impressive on their own. Together they take back most of a day a week.

Three automations I actually run Three automations, each shown as a trigger, a name and the thing it produces. Keyword research runs weekly and returns a scored list of topics. Content radar runs daily and returns a draft queue, not a newsletter. Daily brief runs at 7am and returns one page before coffee. Closing line: the automation drafts, I decide. WEEKLY Keyword research A scored list of topics DAILY Content radar A draft queue, not a newsletter 7AM Daily brief One page before coffee The automation drafts. I decide.
Three triggers, three outputs. The judgment call sits in the middle of each one.

Keyword research, weekly. It pulls search volume and difficulty from an SEO data API, compares it against what we already rank for, and keeps only the gaps. What lands in Notion is a scored list of topics, not a spreadsheet of 4,000 keywords. The scoring is the work. Anyone can pull the data.

Content radar, daily. Industry news, newsletters and research posts, filtered down to the two or three beats I write about. It writes the angle and keeps the link to the primary source attached. What lands is a draft queue, not another newsletter I have to read.

Daily brief, 7am. It reads the overnight email, matches it against today's calendar and my open Notion tasks, and sends one page to my inbox. I read it before coffee. It replaced the twenty minutes I used to spend deciding what today was.

The shape is identical in all three: a trigger, some sources, a first pass at what matters, and one output I read.

Gathering has been getting easier for a decade. What is new is that the machine can now take a first pass at sorting the pile, which used to be an hour of my morning. It is a draft of a decision. The decision is still mine, and it always was.

Nobody automates their life. Stop selling that.

There is a whole genre of content right now promising you can automate your life, set it once and walk away. It is a good pitch. It is not what happens.

An automation is a system, and a system needs an architecture. You decide what it looks at, what it ignores, what a good output looks like, and what it should do when something is missing. Then you run it beside yourself for a couple of weeks, comparing what it produced against what you would have produced. Then you fix the instructions where it went sideways. Only then does it get to run on its own, and even then you read the output.

Mine took real work to get right. Here is what went wrong along the way.

The radar handed me a source that did not say what the summary said. A citation that misstated its own primary source. I now trace every claim back to the original before it goes anywhere with my name on it. The automation finds things. It does not verify them.

Scheduled runs act without asking. A Claude Code routine runs autonomously, which means no permission prompts during the run. Anthropic's own docs say it plainly: every connector you include is usable, including writes, without asking. The default is to include all of them. Strip that list down to what the job actually needs.

A green run does not mean it worked. Also from the docs: green means the session started and exited without an infrastructure error. Whether it did the thing you asked is a separate question, and you find that out by reading the run.

So no, this is not plug and play. What it is, honestly, is this: automation eats the repetitive third of your work. The judgment, the relationships, the decision about what is worth doing at all, that stays yours. If it did not, the output would not be worth reading.

If you want to start this week

Do not start with the impressive one. Start with the thing you do every Monday that you resent.

Write it out as if you were explaining it to a new hire: what to look at, what to ignore, what good looks like, what to do when something is missing. That paragraph is the automation. Run it by hand a few times and fix the instructions where it drifts.

Then schedule it, and pick the location by whether it needs your files or needs your laptop closed.

Two weeks in you will have the honest answer, which is whether you still read the output. If you stopped reading it, it was never worth automating.

The part I am still working out

The old skill was building the branches. That skill is worth much less than it was two years ago, and I spent years getting good at it.

The new one is writing instructions clear enough that something else can act on them, and then checking the work. It looks a lot more like managing than like building.

I am not sure yet whether I am better at this than I was at the old thing.

What did you automate that you ended up turning off? That answer is usually more useful than the ones that worked.

Common questions

What is the difference between Claude Cowork and Claude Code? Both run on the same agentic architecture. Cowork is the workspace inside the Claude app for people who do not write code: it works on your files, your browser and your connected tools, and returns documents. Claude Code is the terminal and desktop version aimed at software: repos, servers, scripts. If you are not going to open a terminal, you want Cowork.

Can Claude run automations while my computer is off? Yes, if you put them in the cloud. Cowork's /schedule tasks run in the cloud, and Claude Code Routines run on Anthropic's cloud infrastructure. The local options, Claude Code Desktop scheduled tasks and /loop, need your machine on.

How often can a Claude automation run? Cloud Routines have a minimum interval of one hour. Local scheduled tasks and /loop go down to one minute. /loop also expires after seven days and stops when the session ends.

Do I still need Zapier, Make or n8n? Yes, and they are not competing with Claude anymore. n8n, Make and Apify all publish MCP servers, so an agent can call your existing scenarios as tools. Keep the deterministic steps in the deterministic tool, where a broken run is loud and easy to find, and let the agent handle the step that needs a judgment call. Most useful automations are a mix of the two.

Can I automate my whole job? No. You can automate the repetitive part, which in most knowledge work is a real chunk of the week. The judgment and the relationships are what is left, and those are the parts that were worth paying for anyway.

What is the biggest risk with agentic automations? Not that they break. That they act. A scheduled run happens with no approval prompts and can write through any connector attached to it. Scope the access to the job, and read the runs.

Sources