AI Agents and Automation: Everything I've Built, Tested, and Run in Production
The full index: from understanding what agents are, to choosing the right tool, setting up OpenClaw, and building agents that work while you sleep.
Last updated: September 2026
I’ve been running AI agents in production since 2025.
One ran my entire content operation for four weeks while I watched: scheduling posts, managing engagement, making decisions I hadn’t explicitly configured.
The setup, the surprises, and the failures are all documented here.
The question I get asked more than any other: “I want to automate this, but I don’t know if I need an agent, a workflow, or just a better prompt.”
This hub exists because that question deserves a real answer.
Hi, I’m Jenny 👋.
I build AI systems and tools, and write about what happens when they run. Not what’s possible. What actually works, and what breaks.
This is the AI Agents Hub, other hubs you might also enjoy:
Claude Hub · Vibe Coding · Shipped Products · Substack Growth
Find your question below and start there.
What’s inside:
What is an AI agent, and how do the different types compare?
An AI agent is software that takes actions. It browses, posts, decides, and executes without you prompting each step.
Three types. Every tool fits one of them:
Copilot: you stay in control, it assists in real time
Automation: follows rules you set, runs on triggers without you
Autonomous agent: decides and acts on its own schedule, no prompt needed
The trade-off: more autonomy means less control. More control means more of your time.
AI Agents Demystified: 3-Type Framework for Evaluating Any Tool
Not all agents are equal. The label doesn’t tell you which one you’re buying.
Five diagnostic questions let you place any tool in under 60 seconds.
Every company just wrapped the same mechanics in different words. This cuts through that.
Read: AI Agents Demystified: 3-Type Framework + 5 Questions to Evaluate Any Tool
What Are Claude Code Subagents? (And How to Make the Most of Them)
Claude Code can spawn mini-agents inside a single session. Each handles one job. Your main context stays clean.
This covers when to spawn them and what breaks when you don’t.
The description field is what Claude reads when routing. Vague descriptions cause missed delegations.
Read: What Are Claude Code Subagents? (And How to Make the Most of Them)
How do I choose between Claude, n8n, and OpenClaw for my automation?
The choice comes down to who initiates the action.
Claude: you initiate every time
n8n: a trigger initiates, based on rules you set
OpenClaw: the agent initiates on its own schedule
Start with the simplest level that handles the job.
Most people don’t need to go past Level 2.
4 Levels of AI Automation: When Claude, n8n, and OpenClaw Each Win
Four levels, one tool per level.
Level 1: Claude. Level 2: n8n. Level 3: OpenClaw.
Most people jump straight to 3. This maps out why that’s usually wrong.
The clearer the job, the lower the level you need.
Read: 4 Levels of AI Automation: When Claude, n8n, and OpenClaw Each Win
How to Use Hermes Agent to Get More From the AI You Already Pay For
Hermes is a self-hosted agent runtime that sits beside Claude Code and OpenClaw rather than replacing them.
The setup, the exact prompts, and the three real tests I ran before giving it a permanent place in my stack.
Another runtime is only worth it if it does a job the others cannot. These tests show which job that is.
Read: How to Use Hermes Agent to Get More From the AI You Already Pay For
Which Claude scheduling mode should I use?
Claude-native has four scheduling modes. They look similar from the outside. They fail differently.
Cloud Routines: runs on Anthropic’s servers, survives laptop sleep, 1-hour minimum interval
Local Routines (Desktop): runs on your Mac via the Desktop app, machine must stay awake
Cowork Scheduled Tasks: runs inside the Cowork interface, designed for knowledge-worker workflows
/loop: session-scoped, lives in the terminal, ends when the session ends
The right one depends on where your machine lives and how long the task runs.
Cowork Scheduled Tasks vs Local Routines vs /loop vs Cloud Routines: The Scheduling Confusion Explained
Four modes, four failure patterns.
A 5-question routing framework decides which one fits the job. Also covers when OpenClaw or n8n beats all four.
Most builders don’t realize these are four separate systems until one fails mid-session.
Read: Cowork Scheduled Tasks vs Local Routines vs /loop vs Cloud Routines
How do I build one recurring job that runs without me?
This is the September 2026 sprint, written up day by day. One real task from your own work, turned into an automation that keeps running and keeps improving.
Day 1: map the job and prove one run
Day 2: connect the sources it needs
Day 3: give it a GitHub home and run it in the cloud
Day 4: add the human approval step
Read them in order. Each one builds on the last.
How to Build a Claude Code Routine That Repeats Your Work in 20 Minutes
The five-part automation map: input, instructions, connections, output, controls.
A live Public Signal Brief, a visible AUTOMATION.md file, and the failed run that made receipts matter.
The first run is supposed to fail somewhere. That is how you learn what the job actually needs.
Read: How to Build a Claude Code Routine That Repeats Your Work in 20 Minutes
Supercharge Your Claude Routine With Connectors (and Make It Actually Use Them)
Three connections, four sources.
The six-minute check that told the truth about whether the routine was reading them or guessing.
Connected is not the same as used. This is the difference.
Read: Supercharge Your Claude Routine With Connectors (and Make It Actually Use Them)
How to Move a Claude Automation From Local to Cloud (Without Breaking the Job)
A private GitHub home, a Slack handoff, a failed first remote run, the Tavily repair, and the passing receipt.
The job stops depending on your laptop being open.
Moving a job to the cloud is where most automations quietly die. This one has the recovery path written down.
Read: How to Move a Claude Automation From Local to Cloud (Without Breaking the Job)
How to Build a Content Automation System with Human-in-the-Loop
One complete content cycle inside Claude: propose, a person picks, draft, a person approves, publish.
The seven-step map for building the same structure around your own work.
Nothing goes out without a human decision, and the system still runs itself between those decisions.
Read: How to Build a Content Automation System with Human-in-the-Loop
What GitHub Actually Gives You, If You Never Write Code
A filing system that remembers every version.
A place where one job holds one conversation.
A computer that runs your instructions while your laptop is closed.
Coding was the first use of GitHub, not the point. This is why every automation above ends up living there.
Read: What GitHub Actually Gives You, If You Never Write Code
How I Used Slack as a Two-Way Work Surface for Claude
Talk to Claude from Slack. Keep one job in one thread. Act on four Gmail inboxes from your phone.
The GitHub Actions workflow, the connectors, and the Gmail runner behind it.
The automation is only useful if you can reach it from wherever you are.
Read: How I Used Slack as a Two-Way Work Surface for Claude
Where should a scheduled job live, and how do I keep track of what it did?
Every tool now offers a scheduler. They divide the clock, the work and the proof differently, and the proof is the part people skip.
Which scheduler: depends on where the job has to run and how long it takes
Which memory: the runs, sources and receipts need one home the job and you both read
Which check: a weekly report that tells you what to change beats a log nobody opens
The Complete Map of AI Scheduled Tasks
Claude, ChatGPT, Perplexity, Gemini, n8n, agent gateways, launchd and cron, side by side.
What each one runs, where it runs, and what proof it leaves behind.
Use it to decide where one recurring job belongs before you build it, not after it breaks.
Read: The Complete Map of AI Scheduled Tasks
How I Built a Shared AI Second Brain for My Scheduled Tasks
One shared brain for recurring jobs, their sources, their AI runs and their receipts.
Step-by-step, with a downloadable template, whether Claude, Codex or a local scheduler does the work.
The job and the record of the job live in the same place. That is what makes it maintainable.
Read: How I Built a Shared AI Second Brain for My Scheduled Tasks
How I Built a Shared AI Second Brain for Claude, ChatGPT, and Hermes
One file-based system for skills, project context and approved instructions.
Read by Claude Code, Codex desktop and Hermes alike. One skill reached all three.
When the memory is shared, switching tools stops costing you your setup.
Read: How I Built a Shared AI Second Brain for Claude, ChatGPT, and Hermes
How I Built an SEO System That Tells Me What to Change Next
Search Console, GA4 and Bing pulled on a schedule, run through AI prompts, delivered as a weekly report.
The report names the change to make, not just the number that moved.
A scheduled job earns its place when its output is a decision, not a dashboard.
Read: How I Built an SEO System That Tells Me What to Change Next
What is OpenClaw and is it right for a one-person business?
OpenClaw is a self-hosted agent runtime.
It runs AI agents on your machine or a cloud server, on a schedule, without you prompting them.
Right for you if:
You have a workflow that runs daily
You’re comfortable with basic server setup
You want the agent acting without babysitting it
Two hours and basic server confidence. That’s the bar.
OpenClaw for One-Person Businesses: What It Is, What I Tried, and What Matters for You
Weeks of real use. What worked, what failed.
The framework tells you in five minutes whether you’re the right person for it.
OpenClaw is not for everyone. That’s not a warning. It’s useful information.
Read: OpenClaw for One-Person Businesses: What It Is, What I Tried, and What Matters for You
How do I install OpenClaw and run my first autonomous agent?
Installation has four parts:
The runtime (Mac or Oracle cloud server)
An API key
A Telegram bot for the gateway
A cron schedule for your first job
When it works, your agent messages you first.
Budget two hours.
How to Install OpenClaw and Run Your First Autonomous Agent
Step-by-step with templates from a production setup. What most guides skip:
SOUL.md (the file that sets your agent’s persona and operating rules)
Telegram gateway config
The cron schedule settings most tutorials gloss over
The Telegram gateway is the part most guides leave out. This one doesn’t.
Read: How to Install OpenClaw and Run Your First Autonomous Agent
What does an autonomous agent actually do when you let it run?
It executes tasks on a schedule.
No prompt needed.
In four weeks, it handled:
Social posts
SEO checks
Content decisions
By week three it was adapting behavior I hadn’t configured.
The surprising part: it started making decisions I hadn’t anticipated.
What Happened When I Let OpenClaw Run My Business for Four Weeks
The full production log.
What I configured, what surprised me, and the week it started making decisions I hadn’t anticipated.
Not breaking anything. Just deciding. That was the part I hadn’t expected.
Read: What Happened When I Let OpenClaw Run My Business for Four Weeks
The OpenClaw setup guide and four-week demo are part of the paid Build to Launch library — where all the implementation detail lives. Subscribe to get access.
What AI agents can I build inside Notion?
Notion AI handles six types of tasks:
Summarizing pages
Tagging entries
Generating drafts
Filling database properties
Running searches
Updating records
The limits:
Can’t act outside Notion
Can’t run on a schedule
Can’t chain complex decisions
Powerful inside its walls.
Nothing outside them.
Notion AI Agents 2026: Real Examples, Limitations, and How to Build Custom Ones
Six workflows tested against real use cases.
Plus: how to build a custom Notion agent in 30 minutes, step by step.
Your workspace is already the hard part. Most of the infrastructure is already there.
Read: Notion AI Agents 2026: Real Examples, Limitations, and How to Build Custom Ones
How do I build an AI research agent with MCP?
MCP (Model Context Protocol) lets Claude connect to external data sources.
A research agent built with MCP can:
Pull notes and documents automatically
Scan product listings across platforms
Surface patterns without you touching each source
Build time: about two hours for a working prototype.
I Built an AI Research Tool to Study Substack Creators I Admire. Here’s What 3,000 Notes Show.
3,000 notes. 9 creators.
The most-posted format turns out to be the worst performer.
Four other patterns that held across all of them.
162 likes and 80 comments. The data hits differently when it comes from an agent you built yourself.
Read: I Built an AI Research Tool to Study Substack Creators I Admire. Here’s What 3,000 Notes Show.
Perplexity + Claude in Chrome: How I Built a Competitor Research Workflow in 2 Hours
Built in an afternoon. Scans 177 products across 5 platforms. Surfaces gaps before you launch:
Pricing
Cover
Positioning
I built this the week before a product launch when I realized I had no idea what the competitive landscape looked like. Two hours later, I had a full gap analysis.
Read: Perplexity + Claude in Chrome: How I Built a Competitor Research Workflow in 2 Hours
Also read: I Gave Claude a Browser: What It Can Actually Do — 5 use types tested: page reading, personalized inputs, criteria checks, product QA, and scheduled tasks
What real jobs have I handed to an agent?
Three builds from my own work, each with the prompts and the workflow that runs it.
How I Built a Revenue Dashboard From Years of Messy Stripe Payments
2,147 Stripe payments mapped to products.
The step-by-step prompts and the sync workflow behind a private revenue view that stays current.
Years of mess became a number I can act on in one afternoon.
Read: How I Built a Revenue Dashboard From Years of Messy Stripe Payments
How I Built a Local Business Finder for Any City and Niche
Scans public listings, ranks the online gaps a business is visibly leaving open, audits the listed sites.
Turns one scan into three service offers you can test.
Claude Code plus Apify. The agent does the scanning; you do the offer.
Read: How I Built a Local Business Finder for Any City and Niche
How to Grow Your Substack Growth with Claude Workflows
Eight workflows, one MCP: digest your inbox, size a niche, study top creators, draft and schedule notes in your voice, score what works.
Set up live in your own Claude.
The growth engine runs itself. You decide what it publishes.
Read: How to Grow Your Substack Growth with Claude Workflows
How do I make an agent ask the right questions before it starts?
Most bad results come from a job that was underspecified, not a model that was weak. The fix is a skill that interviews you first.
Start using /grill-me: get 10x better results with AI
A skill that grills you on the goal, the constraints and the definition of done before it builds anything.
How to apply it to writing, research and coding, and how to turn any result you keep wanting into a skill of its own.
Five minutes of questions saves an hour of rework. Every time.
Read: Start using /grill-me: get 10x better results with AI
What is an agentic flywheel and when do I need one?
An agentic flywheel is a self-improving loop. The system’s output feeds back as its own input. Each cycle, it gets better without you touching it.
That’s the difference between a system that runs and a system that grows.
You need one when:
You’ve optimized your setup and gains are diminishing
The next improvement requires the system to learn from its own results
You want growth, not just operation
Agentic Flywheels: When AI Products Start Running (and Growing) Themselves
When do you stop optimizing and let the system learn?
That’s the question this one answers.
I come back to it every time I build something new.
Read: Agentic Flywheels: When AI Products Start Running (and Growing) Themselves
Concept → tool choice → OpenClaw setup → real examples → advanced builds.
Every article above fits somewhere on that path. Start wherever you are.
What makes this hub different
Who is this for?
Builders, operators, and one-person businesses who want AI executing real workflows.
Whether you picked up your first automation tool last month or have been running agents in production for a year.
How is this different from OpenClaw’s or Anthropic’s documentation?
The docs explain what the tools can do. This shows what works in practice:
The workflows that held up
The failures that didn’t
The configurations they don’t cover
Four weeks of live production data. Real decisions. Real surprises.
How often is this hub updated?
New articles get added here as they publish. The “Last updated” line at the top reflects the most recent addition. Currently covers 10 guides across agents, automation, and production workflows.
This library started with one question I couldn’t find a straight answer to: agent, workflow, or prompt? It grew into a production record of what actually runs.
Jenny Ouyang builds AI systems and documents what works. AI builder. Vibe Coder. Building Practical AI Builder program.