AI agents doing real work, written up as runbooks
Real AI-agent workflows, each traced to a public source and turned into a runbook you can copy.
Set up a personal assistant that asks before it acts
A beginner-to-advanced guide walks through setting up a personal agent where risky actions (sending, buying, deleting) need explicit approval in chat.
Split a huge legacy ETL monolith into smaller jobs
A large bank used an autonomous coding agent to migrate a multi-million-line ETL monolith into sub-modules; the vendor reports about 12x engineering-hour efficiency and over 20x cost savings.
Route email leads into the CRM automatically
Lead emails are read by a model, qualified, posted to the right Slack channel and created as records in the CRM, with a log in a spreadsheet.
Ask for a security-focused review on a risky change
Before merging a change to login, payments or file uploads, request a dedicated security review on the pull request and get findings to fix.
Update a financial model and summarise the assumptions
The agent opens an existing spreadsheet model, adds projections with working formulas while keeping the formatting, and writes a summary of the assumptions it used.
Build interactive charts without knowing the frontend language
Data scientists who don't write TypeScript had the agent build whole React apps to visualise model performance, iterating with small prompt tweaks when the first attempt fell short.
Run a small team of sales bots around a chief of staff
A seller sets up one pinned coordinating bot for meeting prep and inbox drafts, a prospecting bot that researches overnight, and one bot per key account that watches calls, emails and Slack.
Build a cheap 'company brain' your team can ask in Discord
A founder built a shared knowledge assistant for a startup that the team queries in Discord, and reports running costs of only a few dollars a month.
Triage inbound leads into Slack alerts and tasks
New form leads are scored by a fast hosted model; hot ones trigger a Slack alert, all are logged to a sheet and follow-ups become tasks.
Get an automatic second review on every pull request
Comment @codex review on a GitHub pull request, or switch on automatic reviews; the agent reads the diff, follows your repository's review rules and posts a review focused on serious problems.
Schedule a weekly metrics report for Monday morning
A finished agent task can be set to repeat, for example generating the same metrics report every Monday morning from your connected sources.
Turn a GitHub issue into a pull request by mentioning the agent
With the GitHub Action installed, writing @claude in an issue or pull-request comment makes the agent analyse the code, implement the change and push commits for review.
Get new apartment listings that fit, and book viewings
The bot filters new listings against your criteria as soon as they appear, emails landlords for viewing times and prepares applications for the flats you pick.
Run a daily SEO improvement loop
A video walkthrough shows an agent running a daily loop: check search data, pick pages to improve, draft changes and report back.
Qualify B2B leads with a human approval step
The workflow qualifies leads with a model and stores them in a database, but sends an approval email to a person before any follow-up goes out.
Plan a meal and fill the grocery cart
The agent plans a menu for a set number of people, finds the ingredients in an online shop and asks for permission before it actually buys anything.
Condense scattered docs into troubleshooting runbooks
A security engineering team has the agent read several documentation sources and write compact markdown runbooks, which then serve as context when debugging real incidents.
Watch competitors overnight and flag only real shifts
The bot checks competitors for new launches and messaging changes, reviews your own site for stale claims, and reports only the shifts that are gaining traction, with suggested edits.
Set up a team assistant inside Slack
A software agency published a setup guide for running a self-hosted agent as a Slack assistant that the whole team can ask questions in channels and threads.
Score leads from several channels in one sheet
Leads from forms, email and other channels flow into one workflow that scores them with a model and writes the score and reason to a shared spreadsheet.
Analyse three competitors and get an editable slide deck
You name three competitors; the agent browses their sites and public information, compares them and delivers an editable presentation with its findings.
Diagnose a cloud outage from dashboard screenshots
When a Kubernetes cluster stopped scheduling pods, an infrastructure team fed screenshots to the agent, which guided them through the cloud console to IP-address exhaustion and gave the commands to fix it.
Get one daily digest from Slack, email, calendar and notes
An always-on assistant scans your work channels and meeting notes and sends a short readout of what is new and relevant to your priorities, each item with its source and a suggested next step.
Handle lead-to-payment for a small shop over Telegram
An open-source skill pack for small businesses lets an owner run the path from enquiry to quote to payment link by chatting with an agent on Telegram.
Sort incoming lead emails and draft fitting replies
A workflow template watches a mailbox, lets a language model label each new email (hot lead, question, not relevant) and prepares a matching reply.
Get briefed on upcoming client meetings with fresh news
The agent reads your calendar, looks up recent news on each client and prepares a briefing (or a short deck) before the meetings.
Let a non-technical team build its own routing tool
A legal team with no developers prototyped a 'phone tree' tool that helps colleagues reach the right lawyer, showing that departments can build small internal tools themselves.
Investigate a slowdown after a release
After a release makes a page slower, the bot goes through dashboards and traces, names the most likely hotspot and writes a short report that separates facts from guesses.
Run a small studio's operations through agent tickets
A small software studio wrote how it runs company operations with an agent, including turning incoming leads into tracked tickets (they mention two dozen) so nothing falls through.
Build a Figma plugin that mass-produces ad layouts
Marketers had the coding agent build a small Figma plugin that finds the ad frames and creates up to 100 variants by swapping headlines and descriptions, replacing hours of copy-paste.
Turn a vague bug report into a reproduction pack
The bot reads a bug report, repeats the steps in a staging environment with a test account and returns exact steps, expected vs. actual behaviour, screenshots and browser details.
Answer first-line support questions with a self-hosted agent
A commenter on Hacker News described replacing a basic helpdesk workflow with an agent that answers routine questions from internal docs and hands harder cases to people.
Find weak ads in a CSV and write new variants
A growth marketing team built a workflow that reads a CSV with hundreds of ads, spots the underperformers and writes replacements within strict character limits, using two specialised sub-agents.
Reconcile weekly expenses and chase missing receipts
Each week the bot builds an expense summary, matches receipts from the finance inbox, flags missing categories or policy breaks and drafts one follow-up per person.
Run agency work with specialised sub-agents
A marketing agency describes splitting work across specialised agents (research, drafting, reporting) coordinated by one main agent, changing how the team runs client accounts.
Get new hires productive in an unfamiliar codebase
New team members point the coding agent at the repository; it reads the project's guidance files, explains how data pipelines depend on each other and which sources feed which dashboards.
Get a weekly ad-budget reallocation proposal
The bot pulls spend and results per campaign, compares them with the monthly budget and the target cost per customer, and proposes where to shift money, with the numbers behind it.
Cut documentation time for physicians
A physician-developer wrote about using an agent as an operations partner for clinical admin; the author reports saving roughly 30–45 minutes per case on documentation.
Build a candidate shortlist with evidence for each match
For an open role, the bot looks for people who meet the must-have criteria, drops anyone already in the applicant system, explains why each person fits and drafts a personal first message.
Take admin off a one-person trades business
A drainlayer described running quotes, invoices, customer records and stock through an agent from the work truck, so evenings aren't spent on admin.
Research target accounts overnight and queue outreach drafts
A bot works through a CRM list at night, scores each company against the ideal customer profile, picks a few contacts per account and leaves email and LinkedIn drafts for a morning review.
Quality-check field work for a pool service business
A route manager at a pool service company handling roughly 400 pools a week has an agent check technicians' visit reports as they arrive and flag weak ones, which they call their biggest money saver.
Monitor a mushroom grow tent through a sensor platform
A hobbyist published an agent skill that talks to a Mycodo sensor setup, so the agent can read humidity and temperature in a grow tent and report problems.
Track the family budget by chatting on WhatsApp
A family logs spending by sending short WhatsApp messages to an agent, which files each expense into categories and answers questions like how much went on groceries this month.
Book recurring gym classes the moment slots open
A developer's agent books the fitness classes they want as soon as booking opens, one of a few small daily jobs it handles.
Build a home-lab inventory in NetBox with an agent
A home-lab writer had a self-hosted agent discover servers, VMs and network gear and record them in NetBox, instead of keeping the inventory by hand.
One new runbook a week
Short, sourced, and practical. Follow @aitakeaways on X and TikTok or subscribe to KI Klar.