Use case · real workflows

Repetitive work is not hard.
Nobody has ever written out the complete rules.

Everyone has three or four tasks they do every week that have gone years without being automated. Not for lack of tooling: drawing out every exception costs more than doing them by hand. An agent needs no such drawing. Here are five workflows with the exact instruction each one is given.

Copy them as they are · adapt to your own work

Why these tasks are still manual

The complete rule costs more than the task

If these tasks were easy to automate they already would be: we have had workflow tools for twenty years. The reason they are still there is always the same, and it is worth naming.

The cost is in the exceptions, not the normal case. The weekly report is trivial until a vendor changes format. Inbox triage is trivial until the message arrives that looks like two things at once. Automating the easy 80% saves nothing if the remaining 20% forces you to review everything anyway.

An agent moves where that cost sits. There is no need to enumerate the exceptions because there are no rules to enumerate: you give it the goal and the criteria, and it looks at the specific case. What used to be forty branches is now three sentences of prose.

And it brings a risk of its own, worth saying out loud. An agent can be wrong in new ways and very confidently. That is why the five workflows below are all written to the same pattern: prepare the work, flag the odd, do not decide. Start there and widen as trust builds.

How to pick the first workflow: not the biggest one, but the one you did by hand the most times this month while feeling a machine should be doing it. You already have that one described in your head, which is half the work.

The five workflows

What each one replaces

None of these are demos: they are tasks somebody was doing by hand every week. The times are estimates of manual work, not promises.

~40 min/week

The Monday report

Gathering numbers from three or four places, comparing with last week and writing what changed and why it matters. The tedious part is not the summary: it is opening the four dashboards.

~20 min/day

Inbox triage

Sorting what came in, drafting the predictable replies and flagging only what needs your judgement. Nothing is sent without you seeing it.

~30 min/week

Competitor monitoring

Going through the blogs, changelogs and social accounts of three competitors, and telling you only what changed. A rule-based flow breaks here every time someone redesigns their site.

~1 h/week

Preparing posts

Turning one idea into drafts adapted to each network, with their formats and lengths, ready for you to approve or bin.

~15 min/PR

First pass on PRs

Reading the change, checking there are tests, checking it in the browser when it touches the interface, and leaving the obvious comments before a person looks at it.

The shared pattern

Prepare, do not decide

All five leave the work ready and the decision with you. That is what makes an agent mistake cheap: it never reaches the world without somebody seeing it.

Before setting anything up

Which workflow belongs to which tool

An agent is not a better tool than a rule-based automation: it is better for a different kind of task. Choosing wrong costs time in both directions.

What your task looks like Use this Why
Fixed steps, no exceptions, high volume Rule-based automation (Zapier, n8n) Cheaper, faster and auditable step by step
Fixed steps, but the tool has no API Agent with a browser It uses the web interface the way you would; no integration needed
It changes with the specific case Agent There are no rules to enumerate: there is judgement to apply
Something has to be read and weighed Agent This is exactly what a language model does well
It must come out identical every time Rule-based automation An agent is variable by design — here that is a defect, not a virtue
It must answer in seconds, always on Managed service A local agent needs your machine powered on
It touches data that cannot leave your network Local agent Runs on your machine; only inference leaves, or nothing with a local model

With the literal instruction

The five workflows, exactly as they are asked for

1

Monday report — "tell me only what changed"

Schedule a weekly agent that opens your dashboards, compares with last week and writes the summary. The part of the instruction that matters is the last line: without it you end up reading a three-page report every Monday. Instruction: "Every Monday 8:00 — open the sales dashboard, the support one and Analytics. Compare with last week. Tell me ONLY what changed meaningfully and why it matters to me. If nothing changed, reply in one line."

2

Inbox triage — "prepare, do not send"

The agent sorts what arrived, drafts the predictable replies and flags what needs your judgement. Nothing goes out without your approval, which is what makes this workflow safe from day one. Instruction: "Go through the unread mail. Sort into: you reply / I need to decide / ignore. For the first group leave a DRAFT, do not send it. For the second, tell me in one line what has to be decided."

3

Competitor monitoring — where rules break

This is the workflow that best shows the difference: a scraper with CSS selectors breaks every time somebody redesigns their blog; an agent reads it anyway because it is looking at it. Instruction: "Every Monday 9:00 — check the blog and changelog of the three competitors and their X accounts. Compare with last week. Tell me only what is new and what it means for us. If anything is a pricing change, lead with that."

4

Posts — one message, five formats

The agent adapts the idea to each network respecting their limits and conventions, and leaves everything as drafts. What gets automated is the mechanical reformatting, not the decision of what to say. Instruction: "Take this idea and prepare drafts for X, LinkedIn and Instagram, respecting the tone and length of each. Do not publish anything. Flag which one you think works best and why."

5

PR review — the first pass, not the last

The agent does the obvious before a person looks: whether a test is missing, whether the change touches something out of scope, and if it is UI, opening it in the browser and actually checking. Human review still happens, it just starts further up. Instruction: "Review this PR: check there are tests for what changed and that it touches nothing outside its scope. If it touches the interface, open it in the browser and check it. Leave the obvious comments. Do NOT approve it."

The rule that makes this work

Always tell it what to do when it does not know

If you look closely, the five workflows above share one sentence: what to do when things are unclear. That is not a stylistic detail — it is what separates a useful agent from one that makes you check everything twice.

The expensive agent failure is not being wrong: it is inventing something plausible rather than admitting it found nothing. An instruction that explicitly says "if you find nothing solid, say so and stop" removes most of that risk, and costs eight words.

The second rule is just as cheap: state the goal, not the steps. The moment you dictate the steps you are paying for a language model to act as a rule engine, and a less reliable one at that. The value appears when you give it the result you want and the criteria for judging it.

# The skeleton that works across all five workflows WHEN Every Monday at 9:00 WHAT Check X, Y and Z and compare with last week CRITERIA Tell me only what is relevant and why it matters LIMIT Do NOT decide or send anything: prepare and flag ON FAIL If you find nothing solid, say so and stop

Frequently asked questions

Automating work with AI agents

Which tasks should I automate with an agent first?

The ones that prepare work for you to decide on: drafts, summaries, comparisons, first-pass reviews. They have the best value-to-risk ratio, because a mistake never reaches the world without a person seeing it. Leave anything that takes irreversible actions for later, and start with the task you did by hand the most times this month.

How is this different from automating with Zapier or n8n?

In who picks the steps. In Zapier or n8n you write them up front, which is why the flow comes out identical every time — perfect for stable, high-volume processes. With an agent you give the goal and it looks at the specific case, which is what you need when there are exceptions or when the tool on the other side has no API. They do not compete: they cover different halves of the problem.

Do I need to know how to code?

Not for the workflows on this page: they are asked for in prose, like the examples above. What you do need is to describe the result you want and the criteria for judging it, which turns out to be the genuinely hard part. Coding helps when you want to chain several agents or connect your own tools over MCP.

What happens when the agent gets it wrong?

That is why all five workflows are written to prepare rather than to decide. The three measures that actually change the outcome are minimal permissions, results a person reviews before they ship, and running in an environment where you can see and undo what it did. With that, a mistake costs a minute of review instead of an incident.

Can I leave them running while I am away?

Yes, with agent cron jobs that fire at whatever time you set. The honest limitation is that they run on your machine: if it is off, they do not run. For a morning report or an overnight review it works well; for something that must answer at three in the morning no matter what, you need a dedicated machine left on.

What does running these workflows cost?

The platform is a flat capacity price — Free or Ultra at $8 per month — and you pay inference straight to your AI provider with your own key. The real cost depends on how much context each task re-reads, so the way to control it is scoping the request well, not hunting for a cheaper plan.

Does this work for non-engineering work?

Yes, and in fact three of the five workflows on this page never touch code. Research, content, ops and monitoring work just as well; the only requirement is that the task can be described by its result. If you could explain to a new colleague what you want and how you will know it went well, it can be delegated.

Start with the task you did by hand the most

Free forever on Linux, Windows and macOS. Runs on your machine, with your own AI keys.

No credit card · BYOK