◆ AI · automate

Automate a task with AI agents

Chain a few AI agents to do the boring multi-step job while you watch — until one of them confidently does the wrong thing.

4heights
8tasks
8roles do it

AI prompts

do it · improve it · decide · become
AExecute — “help me do it”Act as an automation planner. I want to automate this task: [describe the task, the tools…+
Act as an automation planner. I want to automate this task: [describe the task, the tools involved, and what 'done' looks like]. 1) List every step a human currently does, in order. 2) Mark which steps an AI agent can do vs which need a human check. 3) Pick the simplest tool for the job (a script, a browser agent like Skyvern, or a coding agent like Copilot Agent Mode) and say why. 4) Write the actual first step as runnable code or a concrete agent instruction. 5) Add one 'stop and ask me' checkpoint before anything irreversible. 6) Tell me how I'd test it on fake data first. 7) Flag the one step most likely to silently fail.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed.
BImprove — “do it better”Here's my agent setup / script for automating [task]: [paste it]. Don't rewrite it wholesale.…+
Here's my agent setup / script for automating [task]: [paste it]. Don't rewrite it wholesale. 1) Point out where it'll break if the input is slightly different than expected. 2) Show me where it does something irreversible without a confirmation. 3) Find the step where the agent could loop forever or burn tokens. 4) Add error handling only where it actually matters, not everywhere. 5) Tell me the one thing to log so I can debug it at 2am.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed.
CDecide — “help me choose”I need to automate [task]. Help me choose between: (a) a plain script I write and cron, (b) a…+
I need to automate [task]. Help me choose between: (a) a plain script I write and cron, (b) a browser agent like Skyvern for a site with no API, (c) a coding agent like Copilot Agent Mode inside my repo, (d) a multi-agent setup with AutoGen. For each: rate setup effort, how much it breaks when things change, cost to run monthly, and how easy it is to hand to a coworker. End with a one-line recommendation for MY case and the strongest reason NOT to pick it.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed.
DBecome — “help me grow”Teach me to design AI automations well, using [my task] as the worked example. 1) Walk me…+
Teach me to design AI automations well, using [my task] as the worked example. 1) Walk me through how you'd break the task into agent-sized steps and explain the thinking. 2) Show me the rule for deciding what a human must approve. 3) Point out the two mistakes beginners make with agents (over-automating, no logging) using my example. 4) Give me a checklist I can reuse for the next task. 5) End with one small automation I should build first to practice.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed.

The real tasks

8 of them
I have to run the same data pull and report every Monday.
I have to fix bugs across ten files without breaking things.
I have to scrape prices off a site with no API.
I have to onboard new hires with the same twelve steps.
I have to write and run tests for a whole module.
I have to triage support tickets before a human sees them.
I have to move data between three tools that don't talk.
I have to research fifty competitors and log what they charge.

Who does this

8 roles
Software DeveloperDevOps EngineerData AnalystOperations ManagerQA EngineerStartup FounderMarketing CoordinatorCustomer Support Lead

Questions people actually ask

with the jobs and tasks they touch

Don't hand it the whole job on day one. Give it one clear step, watch it, then chain the next step once you trust the first.

  1. Write the task as numbered steps a human would do, no skipping.
  2. Pick one tool: script, Skyvern for browsers, or Copilot Agent Mode in a repo.
  3. Let the agent do step one only, on fake or copied data.
  4. Add a human 'confirm before doing X' gate before anything you can't undo.
  5. Run it end to end once while watching, then let it go unattended.
see alsoSoftware Developer →

Agents don't know when to quit. They'll retry the same broken step until your token bill looks like a phone number. You set the limits, not the model.

  1. Set a hard max on steps or retries in the agent config.
  2. Add a cost/token ceiling that stops the run cold.
  3. Make each step check 'did this actually change anything?' before repeating.
  4. Log every action so you can see where it got stuck.
  5. Run on a small budget first and watch the spend live.

This is exactly what browser agents like Skyvern are for — they read the page like a person and click through it, so a redesign doesn't instantly kill you the way a brittle CSS selector would.

  1. Describe the site and the exact data you want in plain English.
  2. Give Skyvern the URL and the goal, not step-by-step clicks.
  3. Run it on two or three pages and check the output by hand.
  4. Store results as it goes so a crash doesn't lose everything.
  5. Add a slow-down delay so you don't get IP-banned.

Copilot Agent Mode can read your codebase and edit multiple files, but it works best when you give it one concrete goal and let it show you a diff before it commits.

  1. Open the repo and describe the bug and how you'd know it's fixed.
  2. Tell the agent to find every place the bug touches, not just one file.
  3. Ask it to propose a diff before changing anything.
  4. Run the existing tests, or ask it to write them first.
  5. Review the diff line by line before you merge — always.
see alsoSoftware Developer →QA Engineer →

Most 'automation' is really just moving data from tool A to tool B in the right shape. An agent can do the shaping, but you still need the plumbing — an API key, a webhook, or a browser step.

  1. List the three tools and what data has to move between them.
  2. Find each tool's API or, if there's none, use a browser agent.
  3. Have the agent transform the data into the target's format.
  4. Test with one record and confirm it landed correctly.
  5. Add a check that flags records that failed to move.

A chatbot answers. An agent acts — it takes a goal, breaks it into steps, and uses tools to get there without you typing every move.

  • Chatbot: you ask, it replies, done.
  • Agent: you give a goal, it plans steps and runs them.
  • Agents use tools — browsers, code, files, APIs.
  • Agents can loop: try, check, retry.
  • Multi-agent setups (like AutoGen) have agents check each other.
  • More power means more ways to quietly go wrong.

If the steps never change and there's an API, a plain script is cheaper, faster, and won't hallucinate. Agents earn their keep when the task is messy, the page changes, or the decision needs judgment.

  • Fixed steps + API = write a script, skip the agent.
  • Site changes often or has no API = agent wins.
  • Task needs judgment each time = agent.
  • High volume, low variation = script is cheaper.
  • You need it auditable and predictable = script.
  • One-off exploratory task = agent's fine.
see alsoSoftware Developer →DevOps Engineer →

The idea behind tools like AutoGen is that one agent writes, another reviews, and they argue until the answer holds up — like a coder and a reviewer instead of one person marking their own homework.

In practice it's slower and pricier, so only reach for it when a single agent keeps making the same mistake unchecked.

Almost always it's an input you didn't expect: a blank field, a weird date format, a page that loaded differently. Test data is polite; real data is not.

Feed it your ugliest real examples early, and log what it saw right before it broke.

Start by writing the exact steps you do now in plain English, then paste that into a coding agent and ask it to build the smallest version first. You don't have to code — you have to describe clearly.

Get one report running by hand-triggering it before you schedule it to run alone.

Never let an agent do anything irreversible without a human 'yes' — deletes, sends, payments, publishes. Put a hard confirmation gate in front of every one of those.

And keep backups, because 'the agent did it' is not a recovery plan.

Different jobs. Copilot Agent Mode lives in your code. Skyvern lives in the browser. Pick by where the work actually happens.

JobCopilot Agent ModeSkyvern
Edit code across filesBuilt for itNo
Click through a websiteNoBuilt for it
Site has no APINot the toolYes
Runs inside your repoYesNo
see alsoSoftware Developer →Data Analyst →

Only if a lone agent keeps shipping mistakes nobody catches. Multi-agent buys you a built-in reviewer at the cost of speed and money.

FactorSingle agentAutoGen (multi)
Setup effortLowHigh
Cost per runLowerHigher
Catches own errorsRarelyBetter
SpeedFasterSlower
see alsoStartup Founder →Software Developer →

They'll eat the repetitive, no-judgment parts — the copy-paste, the same report, the ticket sorting. That's real, and pretending otherwise is dumb.

But someone has to design the automation, catch it when it lies, and decide what's worth automating. The person who understands the whole task and can wire up the agent is worth more, not less. Be that person.

Trust the parts you've watched succeed a dozen times on real data, and never trust the destructive parts without a gate. It will fail eventually — the question is whether it fails loud or silent.

Set it up so failure is loud: logs, alerts, and a hard stop before anything you can't undo. Then let it run, and check the logs like you'd check a smoke alarm — regularly, not never.