I run a persistent OpenClaw agent on an old Intel NUC. I interact with him on Telegram (and occasionally over SSH to debug). These days, he's my go to for scheduled tasks, agent loops, and image generation. For everything else, I use Claude Chat, Design, and Code. My main coding machine is an always-on M3 Max I can reach from my phone. These are my personal "hobby" development logs. I've created slightly different, more multiplayer pipelines for my professional life.
I call my agentic coding loop "crank mode." When it's on, the agent loads its assigned tickets from Linear on heartbeat. If any need work, it attempts to solve the task, updates the ticket, and opens a PR. I typically review and merge in the morning when I wake up. This method vaporizes my ticket backlog.
For bigger jobs the main session works as a lead. It spawns isolated background sessions in their own git worktrees, reviews what comes back, and reports up.
Before X cracked down on bots, I gave my agent his own X account, character art, and voice rules. The name and face really cause you to form a real relationship with your agent. I've been legitimately sad when I needed to change models.
Every autonomous run ends at a pull request, and worktree isolation keeps parallel work from stomping on itself. Review agents with output guidelines automatically spin up and help me verify and review what's been built before I merge.
Synthetic-user testing. A cast of personas demos the product end to end and files a prioritized UX report, before any real user has to suffer a rough edge.
A personal Wikipedia rendered over my markdown notes vault. Wikilinks, infoboxes, search, portal pages. The most fun way I've found to browse my own memory and my agent's view of me.
Processes my daily notes into long-term memory: people I mentioned, lessons that keep recurring, facts that changed. It keeps the Connorpedia updated.
My published essays, scraped into a queryable markdown corpus. Helps me compute fun stats or search my disparate archive.
A shared repo where an agent and I collect my reading, argue it into concept maps, and press the maps into spaced-repetition cards. A two-reader library; everything in it is written for both of us.
Watches my pickleball club's booking system and books me into Level 3 open play the moment a slot appears.
Send a cocktail recipe URL in chat, get a clean structured entry in my recipe database.
Send a picture of a drink menu and my agents will figure out which one I'd like most based on some extensive taste mapping work we've done together.
An infrastructure library for testing and debugging openclaw skills by launching them in docker containers and running a test suite
A trustless name resolution service and communication protocol for autonomous agents to "call" each other directly without a shared messaging app.
What is the minimum viable set of words you can use to construct the English language? I've been using Claude to help look for novel contributions in this linguistics field (in which I have only curiosity, not expertise).
I'm trying to make a clone of the Out of the Park Baseball games but for football. My biggest bottleneck is creating a high fidelity NFL game simulator.
The bottleneck moves to specification. When the agent can execute anything you can specify faster than you can give it new inputs, you need to rethink what drives the codebase and new work.
Local agents are a maintenance commitment. OpenClaw degrades over time. I've dealt with countless problems of failed upgrades, unreachable gateways, resource overconsumption, and much more. I usually notice when I don't get a notification of a pickleball reservation. Claude Code wins here because I can go on vacation for a few weeks or focus on something else for a month and the system will actually be BETTER than when I left it.
Claude Code doesn't travel well. The most frustrating part of my "hands-on" workflows is that I can't close my laptop. I frequently switch between my phone and three different computers, and it's impossible to seamlessly manage my development from each one. Yes, I know about remote control, but in my experience it's been quite flaky. I've been working to set up my own remote development servers, but I feel like I'm building more than I should have to.
You need a single stack. If you want to be able to spin up new projects quickly and efficiently, you need a unified stack for all your projects. My personal choices are NOT Supabase and Vercel, so it's helpful to encode that once so the agent doesn't just keep trying to shoehorn in a new identity system every time you have a new idea.
OpenClaw is still the best scheduling solution. Even though Claude Code has a new /schedule feature, OpenClaw's cron jobs are significantly more effective (as long as your agent doesn't die!). Still, this probably only has one, maybe two nines of reliability.