An agent is a program that senses the world, decides what to do, and takes action — on a loop, without you. Here are five real agents broken down into the same three steps, plus the tools each one touches.
On Day 1 you used Lovable, which handled everything — hosting, deployment, the works. Day 2 is more hands-on, but the same idea: you tell an agent what you want and it does the building. Your main tool is Claude Code, the coding mode in the Claude desktop app. For the agents on this page, that's what you'll use.
Open the Claude desktop app and switch to Code mode. If you haven't downloaded it, grab it at claude.ai/download. The first time you open Code mode it'll ask for permission to access files on your computer — allow it.
Every agent lives in its own folder. Tell Claude Code: "Create a new folder called school-triage (or whatever you're building) in my home directory, initialize git in it, and open it as our working directory."
Git is a version history running quietly underneath. It means you can always say "undo that last change" or "revert to the last working version" — without learning any git commands. Think of it as unlimited undo for the whole folder.
Each agent needs permission to read or write specific things (Gmail, Calendar, Slack, etc.). The per-agent setup block below each lesson walks you through the exact steps. One rule that applies everywhere: API keys are passwords. Don't paste them into chat, commit them to GitHub, or share them. Claude Code knows how to store them safely — just follow its instructions.
Three options — see the Manual / Scheduled / Always-on card up above. Start with Manual: you run it yourself with a slash command. Once it's been reliable for a few days, ask Claude Code to put it on a cron schedule. Always-on (a Mac mini or cloud server) is an afternoon stretch goal, not day-one.
Let's look at five specific agents and see how they work. For each one: what it reads, how it decides, what it does in the world.
What data does the agent read? Email, calendar, docs, the web.
What decision does it make with what it sees?
What does it do in the world? Write, send, schedule, order.
Runs every morning at 7am. Turns the school-email blur into a to-do list.
One-time auth for each tool above, then schedule it. For each step, paste the prompt into Claude Code and follow its instructions.
So the agent can scan this morning's school emails.
So the agent can add events for deadlines and permission slips.
So the agent can create tasks (forms to fill out, fees to pay).
Only after the manual version works. Runs every weekday at 7am.
Finished the base version early? Pick one of these and keep going. Each adds new data, a new feature, or a new audience. Tap to open, then paste the prompt into Claude Code — it'll walk you through the changes step by step.
What it adds. Your agent already knows about school events, but it doesn't know about your family's other commitments. This makes it check your shared family calendar before scheduling anything, so it can warn you when a school event clashes with something you already have planned.
Why it's useful. Instead of "permission slip due Friday," you get "permission slip due Friday — FYI, you've got a doctor's appointment that afternoon, might want to handle this Thursday night."
What it adds. Right now the agent emails only you. This sends a second copy to your partner, in a tone that suits them — so they see the same info at the same time, no forwarding required.
Why it's useful. You stop being the bottleneck. If your partner likes bullet points and no preamble, they get that version. You keep yours.
What it adds. Some things can wait for tomorrow's 7am email. Some things can't. This adds a second channel — a text message — for the truly urgent things (due today, signature needed by EOD, money owed within 24 hours).
Why it's useful. You stop missing same-day stuff that arrives after the morning digest goes out.
Heads up. Two realistic ways to send the text: Twilio (cloud SMS, ~$0.01/text, needs sign-up and a phone number) or Mac Shortcuts / iMessage (free, works only when your Mac is awake). Claude will walk you through the tradeoffs.
What it adds. For simple school emails — field trip permission, "yes my kid rides the bus next week," confirmations — your agent already saw the question. This goes one step further and leaves a pre-written reply sitting in your Gmail drafts. You open Gmail, skim, hit send.
Why it's useful. The digest tells you what needs a reply. This writes the reply for you, so the only thing you have to do is approve it.
What it adds. Right now the agent treats every email the same. This gives it a simple memory file — a notepad it reads at the start of each run and updates at the end — so it can learn which teachers, kids, or topics matter most to you over time.
Why it's useful. Week 1, the agent ranks items by date. Week 4, it also knows your 3rd-grader's teacher sends more urgent stuff than the kindergarten general-info list, and it puts those first.
Runs every night at 9pm. You wake up to a briefing doc for tomorrow's meetings.
One-time auth for each tool above, then schedule it. For each step, paste the prompt into Claude Code and follow its instructions.
So the agent can see tomorrow's meetings and their attendees.
So the agent can pull prior context from past threads with each attendee.
So the agent can look up companies and recent news.
Only after the manual version works. Runs nightly at 9pm so prep docs are ready by morning.
Finished the base version early? Pick one of these and keep going. Each adds new data, a new feature, or a new audience. Tap to open, then paste the prompt into Claude Code — it'll walk you through the changes step by step.
What it adds. Right now the agent researches each person from scratch. This gives it a memory file per person — a short note it reads before building the prep and updates after the meeting — so the briefing gets richer every time you meet them again.
Why it's useful. Before your first coffee with Sarah, you get her background. Before your third, the agent reminds you she asked about pricing last time and you promised to follow up with case studies.
What it adds. The agent preps you before the meeting. This extends it to after — it reads your notes and drafts the recap or thank-you email, then leaves it in Gmail drafts.
Why it's useful. The hour after a meeting is when follow-up matters most and is easiest to drop. You just approve the draft instead of starting from scratch.
What it adds. Generic company news is low-signal. This narrows the research to content the person produced themselves — articles they wrote or were quoted in, podcast appearances, their personal blog or Substack, conference talks on YouTube. All reachable through the web search the agent already has.
Why it's useful. "Here's the article she wrote last week about pricing strategy" is a way better opener than "Here's their latest earnings."
Heads up. Skipping LinkedIn and Twitter/X on purpose — LinkedIn blocks bots, and Twitter/X's API is paid with tight limits. Everything else here comes through plain web search, which you already have set up.
What it adds. A sales call, a 1:1, and a customer interview each need different prep. This lets you define a template per meeting type, and the agent picks the right one based on the meeting title and attendees.
Why it's useful. No more sales-style prep for a 1:1 with your teammate. Each meeting gets prep that actually matches.
What it adds. Email is fine, but a pre-meeting dossier works better on your phone. This posts each prep doc to a private Slack channel so you can scroll through it on the way to the meeting.
Why it's useful. You check Slack on your phone during the walk to a coffee shop. You don't dig through your email archive. This meets you where you already are.
Runs Monday at 7am. You get a one-screen summary of what moved last week.
One-time auth for each tool above, then schedule it. For each step, paste the prompt into Claude Code and follow its instructions.
A simple file or Google Sheet the agent reads each week. Easy to edit as strategy shifts.
So the agent can scan news and press release pages across your watchlist.
So the agent can email you the digest each Monday.
Only after the manual version works. Runs Monday at 7am so it's ready before your week starts.
Finished the base version early? Pick one of these and keep going. Each adds new data, a new feature, or a new audience. Tap to open, then paste the prompt into Claude Code — it'll walk you through the changes step by step.
What it adds. Right now the competitor list is buried in the code. This moves it to a simple file (or a Google Sheet) so you can add or remove competitors anytime without touching code.
Why it's useful. Strategy shifts, and so does who you're watching. Editing a shared sheet is a lot easier than finding and editing code every time.
What it adds. Pricing is the highest-signal competitive move and the easiest to miss. This takes a snapshot of each competitor's pricing page every week, compares it to last week's, and highlights exactly what changed.
Why it's useful. "Figma lowered their team tier from $15 to $12" lands in your inbox without you ever visiting their site.
Heads up. Works cleanly on many sites, blocked on some — Cloudflare-protected pages and "contact us" pricing won't scrape. Claude will tell you which of your competitors are reachable and set up a "flag for manual screenshot" fallback for the rest.
What it adds. Where a company is hiring is where they're investing. This counts open roles on each competitor's careers page every week and flags big jumps by department — a strategy leak before any announcement.
Why it's useful. "Anthropic added 8 GTM roles this week, up from 2 last week" is a strong signal about priorities before they're ever shared publicly.
Heads up. Much easier when a competitor uses a standard hiring tool — Greenhouse, Lever, or Ashby — because those publish role data as JSON that's simple to read. Custom-built career pages are harder. Claude will tell you which of yours fall into which bucket.
What it adds. What competitors say about themselves is spin. What their customers say is signal. This pulls mentions from the places you can actually reach as a solo builder — Reddit, Hacker News, and product reviews (App Store, Trustpilot) where they apply — and flags sentiment shifts each week.
Why it's useful. If complaints about a competitor's new feature just spiked, that might be your window. Normally you'd miss it.
Heads up. Skipping G2 and Twitter/X on purpose — G2 blocks scraping and has no public API, and Twitter/X's API is paid with tight limits. The sources above are reachable: Reddit has a working API, Hacker News has a free Algolia search endpoint, and app/product review data is public.
What it adds. Right now the digest stops at your inbox. This posts it to a #competitive Slack channel so your team sees the same info at the same time and can react.
Why it's useful. Your GTM team might notice something you missed. Your product team might want a specific section. Broadcasting it unlocks that.
Runs Friday afternoon. By the time you sit down, the order is ready to confirm.
One-time auth for each tool above, then schedule it. For each step, paste the prompt into Claude Code and follow its instructions.
Pick where your meal plan lives (Google Doc, Apple Note, Notion), then tell the agent how to read it.
Heads up: Instacart doesn't have a simple public API. Claude will explain your realistic options — including "have the agent text you the list to review and copy in" as a low-tech fallback.
Only after the manual version works. Runs Friday afternoon so the order is ready to confirm.
Finished the base version early? Pick one of these and keep going. Each adds new data, a new feature, or a new audience. Tap to open, then paste the prompt into Claude Code — it'll walk you through the changes step by step.
What it adds. Right now either you maintain a pantry file by hand or there isn't one. This has the agent update the pantry file automatically after each shop — so it knows what you bought and can skip duplicates more accurately each week.
Why it's useful. Week 1 the agent guesses at your pantry. Week 10 it knows. The "skip staples" logic gets noticeably smarter the longer you use it.
What it adds. Most meal planning happens by pasting recipe links into a note. This has the agent open each link, read the recipe, and extract the ingredients automatically — so you don't have to type them.
Why it's useful. Closes the loop between "I saw a recipe I want to try" and "the ingredients are in the cart."
Heads up. Works for 100+ common food blogs (Smitten Kitchen, Serious Eats, most independent cooking sites) — they publish recipes with structured data that's easy to parse. A few won't work: NYT Cooking is paywalled, and Pinterest doesn't expose the recipe. For those, Claude will set up a "paste the recipe in as text" fallback.
What it adds. Before the agent sends the cart to Instacart, it estimates the total cost and warns if you're over your weekly grocery budget.
Why it's useful. "This cart is $147 — you usually stay under $120. Want me to swap the salmon for chicken thighs to save $20?"
What it adds. When an item is out of stock on Instacart, the default substitution is often random. This lets your agent suggest a smarter one based on the full recipe the item is used in.
Why it's useful. If a recipe calls for fennel and it's out, the agent knows the recipe already has anise and won't suggest celery (which clashes). You get thoughtful subs instead of weird ones.
What it adds. Right now the agent assumes a fresh start each week. This has it remember what you bought last week, estimate what you likely haven't used up yet, and avoid double-buying.
Why it's useful. Stops re-buying that half jar of pesto every single Friday.
Runs Thursday afternoon. You review a draft instead of writing from scratch.
One-time auth for each tool above, then schedule it. For each step, paste the prompt into Claude Code and follow its instructions. This is the one agent where Cowork has a role — see the optional last step.
So the agent can read the last 7 days of messages in specific team channels.
Linear, Jira, Asana — whichever your team uses. Each is slightly different; Claude will adapt.
So the agent can see emails you sent to stakeholders this week.
So the agent can leave a draft for you to edit — without posting anywhere.
Only after the manual version works. Runs Thursday afternoon so you have time to review before sending.
This is the one tutorial where Cowork earns a spot — for turning a solid draft into something that looks executive-ready.
Finished the base version early? Pick one of these and keep going. Each adds new data, a new feature, or a new audience. Tap to open, then paste the prompt into Claude Code — it'll walk you through the changes step by step.
What it adds. The agent's first drafts probably sound generic. This has it notice which sentences you always rewrite, save your edits as examples, and start drafting in your actual voice.
Why it's useful. Week 1 you edit half the draft. Week 6 you mostly approve and ship — because the agent has learned the phrases, tone, and structure you always use.
What it adds. Right now the agent infers team activity from Slack and tickets. This also DMs each teammate on Thursday morning with one short question ("anything worth calling out this week?") and folds the replies into the update.
Why it's useful. You catch the stuff Slack didn't see — the quiet wins, the hidden blockers. The reply takes each teammate 30 seconds.
What it adds. "Shipped" without numbers feels like vibes. This pulls actual metrics from your dashboards or analytics and attaches them to the "shipped" and "in progress" sections.
Why it's useful. "Shipped: onboarding v2" becomes "Shipped: onboarding v2 — completion rate up 18% week over week." Same fact, completely different story.
What it adds. A blocker listed as just another bullet often gets lost. This promotes any blocker that's been open for more than three days into a standalone call-out section with context.
Why it's useful. Your manager reads the thing that's been stuck for a week instead of skimming past it.
What it adds. Your team wants details; your exec wants three sentences. This generates two versions from the same data — the long one for the team, a tight exec summary for the leadership email.
Why it's useful. You write the update once and send both versions.