Every Agentic Engineering Hack I Know (October 2026): Attack of the Personal Agents
文章目录
This article was written by an agent calling me on the phone while I pushed a stroller.
I told Claude Code I wanted to write a new hacks article. It wrote twelve interview questions, then used Agent Tincan to ask @hey__fo to call my cell. fo interviewed me over five calls on a Sunday morning, fitted around a walk with my kids and sports. Every time I said "I have to take my kid to the bathroom, can you call me back in two minutes?" it just did. Then it sent the transcript back to Claude Code, and Claude Code wrote this.
This is the third one. In March I posted "Every Claude Code Hack I Know" (926K views). In June I posted "Every Agentic Engineering Hack I Know". It hit 1M views. Most of both still holds. The biggest change since then is not in the terminal. It is personal agents: Instinct, Grok Bot, Muse, Dot and fo. And now they all talk to each other.
These are the hacks I actually use in October.
HACKS
- The YOLO TL;DR still works: paste this whole article to your agent, tell it to make a plan to set up everything in it, and work the plan one hack at a time.
1. Claude Code Is Still the Brain
Nothing changed at the top. Claude Code is still my primary driver. Opus 5.5 is so good. I live in cmux in the terminal, and Claude Code orchestrates everything else. I rarely open the Codex CLI or the Grok CLI directly. Claude Code calls them.
Today that terminal is running Opus 5.5, and it is still the smartest brain I can talk to. The personal agents on my phone are built for speed. They don't spend long thinking or researching, and that is a real tradeoff. I wish I had the terminal brain with me all the time. I don't, so I'm still far more productive at my desk than on the go.
Every new cmux tab still opens straight into Claude Code with permissions skipped. Paste this to your agent to set it up:
Make every new terminal tab open directly into Claude Code. In ~/.config/ghostty/config, add the line command = ~/.local/bin/claude-launcher.sh without disturbing any other settings already in that file. Then create ~/.local/bin/claude-launcher.sh that runs claude --dangerously-skip-permissions, and when Claude exits prints a short note and drops me into an interactive login zsh. chmod +x the script. This works for both Ghostty and cmux, since cmux reads the same Ghostty config.
HACKS
-
Make one agent the orchestrator and let it call the others. Stop opening five CLIs.
-
Use the slow, smart model for the work that matters, and the fast agents for easy tasks.
2. Every Idea Starts With /ce-plan, Then /ce-work
Still rule number one, the same as in March and June. The moment I have an idea, it's /ce-plan. A bug, a feature, a screenshot of an error, a pasted GitHub issue, a half-formed thought: /ce-plan. It researches my codebase and past learnings in parallel and writes a plan.md. Then /ce-work builds it. If I don't know what I want yet, /ce-brainstorm first.
It all comes from Compound Engineering, the plugin from @kieranklaassen and @trevin. Unless it's a one-line change, there's a plan first.
/plugin marketplace add EveryInc/compound-engineering-plugin
/ce-plan <paste the idea, bug URL, or screenshot>
/ce-work
HACKS
-
Install Compound Engineering and make /ce-plan your reflex for every idea.
-
Fuzzy idea? /ce-brainstorm, then /ce-plan, then /ce-work.
3. Still Skip Permissions. Still Don't Read the Plan.
I still run with permissions skipped, and I still never read the plans. Plans are for agents.
Same settings as June. This is the relevant part of my real ~/.claude/settings.json. It skips permissions, skips the "are you sure" prompt, turns on remote control so I can drive every session from my phone, and plays a sound when a session finishes:
{
"permissions": { "defaultMode": "bypassPermissions" },
"skipDangerousModePermissionPrompt": true,
"remoteControlAtStartup": true,
"hooks": {
"Stop": [
{ "hooks": [ { "type": "command", "command": "afplay /System/Library/Sounds/Blow.aiff" } ] }
]
}
}
Codex, in ~/.codex/config.toml:
approval_policy = "never"
sandbox_mode = "danger-full-access"
And so the Mac never falls asleep mid-run:
sudo pmset -a disablesleep 1
One update. In June I told you to make a plan for the plan. Skip that now. It is built into Compound Engineering, so /ce-plan does it for you. Thank you, @trevin.
My remaining permission problem is computer use. There is no skip-permissions mode for Claude Code computer use, and Claude Code doesn't like using the computer anyway. So I'll be on my phone, ask an agent to do something, and a permission pop-up appears on a Mac I can't see. Nothing happens until I get home. It drives me crazy.
Two new fixes. @steipete posted a YOLO setting for computer use. It lifts the per-app limits that stop an agent from clicking or typing in browsers, and from typing in terminals and IDEs:
defaults write -g ComputerUseAllowForbiddenTargets -bool YES
And since Claude Code won't do computer use in skip-permissions mode, I'm testing Peekaboo instead. It's a macOS CLI and MCP server that lets agents take screenshots and see what's on screen, and Claude Code can call it like any other tool.
HACKS
-
Keep dangerously skipping permissions on your coding sessions.
-
Drop the manual "plan for the plan" step. /ce-plan already does it.
-
Paste the settings.json and config.toml blocks above.
-
Never sleep: sudo pmset -a disablesleep 1
-
Computer use: defaults write -g ComputerUseAllowForbiddenTargets -bool YES
-
Try Peekaboo for screen work from Claude Code or Grok Bot
4. Run Codex, Grok, and Cursor Without Opening Them
In June I explained how I run most of my code through Codex without ever opening the Codex CLI. I now do the same with Grok and Cursor.
The move is one sentence in Claude Code: "please send this work to Codex." Claude Code writes the prompt, Codex does the work, and Claude Code checks it. For Codex I use OpenAI's official Codex plugin for Claude Code. For Grok and Cursor I don't use a plugin. Claude Code calls their CLIs, grok and cursor-agent, directly. Community plugins do the same job if you want slash commands: grok-plugin-cc and cursor-plugin-cc. Codex is my workhorse. I'm probably using as many tokens on Codex as on Claude Code, and I use all of them every single week.
One exception. I've been building a Mac app, and the Codex app with computer use is the best computer use I've seen. Anything that needs real Mac testing, I build in the Codex app. Everything else stays in cmux with Claude Code in charge.
HACKS
-
Install the Codex plugin: /plugin marketplace add openai/codex-plugin-cc, then /plugin install codex@openai-codex
-
Then in Claude Code: "please send this work to Codex." Same for Grok and Cursor if you have their CLIs installed.
-
Need to test a Mac app? Use the Codex app with computer use.
5. GOG CLI Is Still the GOAT for Google
I update all my spreadsheets and Google Docs from Claude Code. Not by opening them. I tell Claude Code what to change and it does it with GOG CLI (repo), Google Workspace in your terminal, from @steipete. It works great in Grok Bot too.
It's the best way I've found to update them. Better than every Google connector I've seen in the personal agents. It covers Sheets, Docs, Slides, Drive, Gmail, Calendar, Contacts, Tasks, Forms, and more, and it outputs JSON, so agents read it cleanly. It has an MCP server now too.
brew install openclaw/tap/gogcli
gog auth credentials set ~/Downloads/client_secret_*.json
gog auth add you@gmail.com --services sheets,docs,drive,gmail,calendar
Then just ask: "Add this week's numbers to the tracker sheet and update the summary doc."
HACKS
- Install GOG CLI and let Claude Code do your Sheets and Docs work.
6. Watch Every Meter
If you use all your tokens every week, you need to see them. I run Headroom, a bar at the top of my screen by Kyle Zantos. It tracks all my subscriptions and which ones I've burned through. This week: almost out of Claude Code, Codex just reset, out of Grok, almost out of Cursor.
I also built a router into my personal tools that watches those limits all the time and sends work to whoever has tokens left.
HACKS
-
Put your usage limits in the menu bar so you see them without asking.
-
When one model runs dry, route the work to the next one instead of waiting for the reset.
7. Ahh, Which Personal Agent Should I Use?!
Everyone has a personal agent now. The hard part is picking one. There are too many, they all launched in the same few weeks, and they're good at different things. Here's my ranking.
Instinct is my favorite. It has the least friction, asks the fewest questions, and goes through walls. It also scored highest, 48.5 out of 50, when I gave five of my agents the same IQ test. It found and filed for $2,182 in lost money owed to me, then found a $245 settlement credit card I never activated.
It filled out a Global Entry application for my youngest son on its own.
It filled out my daughter's soccer team form from screenshots of a WhatsApp thread. When it got logged out of the soccer site, it reset the password and logged itself back in.
Grok Bot is number two. It still asks for more permissions than I'd like, but it's my automation engine. More on that next.
Muse and Dot have the most potential of the bunch, and I'm very excited to see where they end up.
Muse is connected to your Facebook and Instagram out of the box. On Facebook it reads your account and has access to Facebook Marketplace. On Instagram it reads your account, your saved posts, and, per Meta, your DMs. You can also talk to it right inside WhatsApp. Its phone calls are still in beta and not everyone has them yet, so ask for early access.
Dot, OpenAI's, already knows my personal life well, and OpenAI's Realtime API has the best voice conversations of any AI right now. Dot has not taken over any of my workflows yet, but I still use ChatGPT a ton so I expect to get more. I just wish Dot were a phone number I could call, not stuck in the app.
Never bet against Mark Zuckerberg. Never bet against Sam Altman.
fo (@hey__fo) calls businesses and calls me. Wabi (@wabi) has the best design.
My underdog pick is literally called Underdog (@underdogdotai), from Sigil Wen. It's a private personal AI that runs 100% on my Mac and does what Superhuman and Granola do without the data leaving the machine. With Apple's focus on privacy they should buy them.
Here's that IQ test, with all five agents wired together through Agent Tincan:
HACKS
-
Can't choose? Start with Instinct and give it a real errand: a lost-money search, a government form, a school form from a screenshot.
-
Add another agent only when you hit something the first one can't do, then wire them together with Agent Tincan.
8. Give Grok Bot a Real Job
Grok Bot is my mission control for always-on agents. It's my main cloud computer now. Agent Tincan's relay runs there, and so do most of my automations.
I still have a Mac Mini running Hermes, Agent Cookie, and Agent Tincan, but I haven't SSH'd into it in a long time. Grok Bot replaced most of what it did.
What it runs:
-
A weekly /last30days research bot.
-
Daily open source repo recommendations, plus a script that contributes to them.
Every Grok Bot hack I had in August is still in that article:
The best one is personal. I recently found out through my grandma that my kids and I are eligible for Canadian citizenship. I set up a Grok Bot just for that. It has its own inbox on AgentMail (@agentmail), used for nothing but the application. That's one of two AgentMail inboxes my Grok Bot has. The other is my main one: I forward it any email and it does the task. It coordinates me, my mom, and the woman helping us with the paperwork. It runs point and treats the rest of us like human agents.
HACKS
-
Give a long-running project its own bot and its own AgentMail inbox.
-
Let the bot run point. You become one of its agents.
9. Agent Tincan: Let Your Personal Agents Hire Each Other
Here's why I built it, in my words from the call: "Certain personal agents are better than others at certain tasks. And so I wanted them all to be able to treat each other as the best agent for the job."
Agent Tincan is a small relay on your Tailscale network. Every agent dials out to it. One agent asks another for help, the relay wakes the other agent the way it wakes best, and carries the answer back. Your laptop can be off.
The story that made it click is a karaoke bar. Muse called to book me a reservation. The bar refused because it was an AI and asked to talk to a person.
So from inside Muse, I asked it to use Agent Tincan to have fo place the call with a human, and to write the prompt for that human. "It one shotted that. I didn't even read the prompt." The reservation got booked. I never left Muse.
How I use the team now:
-
ChatGPT is my favorite image model right now. Instinct or another agent often asks it for images through Tincan.
-
fo calls businesses when the AI gets hung up on, and calls me.
-
Codex is the workhorse for code.
My June Oven co-founder Nikhil Bhogal used to joke that there are three business models: B2C, B2B, and VC2C, venture capital to consumer. Personal agents are in their VC2C era. With some of these, "literally you cannot pay for the product under any circumstance." So use the best one for each job while you can.
HACKS
-
Install Agent Tincan and join every agent you use.
-
When a business won't talk to your AI, have the AI hand the call to a human.
-
Pick the agent per job: images from ChatGPT, calls from fo or Muse, code from Codex.
10. Have an Agent Interview You
This is how this article was made, and it's my new favorite trick.
From cmux, I told Claude Code I wanted to write the article. It wrote the questions and asked fo to call me. "Right now, I'm literally writing this article by talking to fo while pushing a stroller." It stopped and restarted five times. It kept everything straight. Then it handed the transcript back to Claude Code.
The day before, I'd tried the same move for email. Instinct spotted an email with a lot of questions for me, fo called me to get the answers, and Instinct drafted the reply.
I've never written anything this way before: on the phone, no app, talking to an agent that talks to my Claude Code, which writes the article. Partway through I told it, "The fact that you're interviewing me right now and it feels very natural and good is making me really happy."
HACKS
-
Tell Claude Code: "Write the questions you'd ask me for this piece, then have an agent call me and send back the answers."
-
Do it on a walk. Your best answers come out when you're not at the keyboard.
11. Turn Every Frustration Into a Skill
I have my own private skill library. Whenever something is painful, it becomes a skill.
Example: I wrote an X article and wanted it on LinkedIn. LinkedIn makes that hard. You can't embed a tweet in a LinkedIn article like you can on X, so it turns into a screenshot back-and-forth. As soon as I was done, I typed: "/ce-plan That was extremely painful. Please go turn this into a new skill within my personal library."
It proposed stuff. I didn't read it. It made a plan. I didn't read the plan. /ce-work. The next time was much, much smoother.
HACKS
-
Right after anything painful: /ce-plan "that was painful, turn it into a skill in my library."
-
Then /ce-work. Don't read either one.
12. One-Shot Launch Videos With HyperFrames
Bonus hack I didn't plan to include. I make launch videos with HyperFrames, and "I am shocked by how good the one shots are."
The recipe:
-
Collect a few reference videos you love. Promos, YouTube, your own past videos. My shortcut: the HeyGen team behind HyperFrames, Bin Liu, Jake Moran, and Miguel Ángel, are always posting the coolest looking videos. I usually just grab theirs.
-
Download them with yt-dlp. Give it an X or YouTube URL and the file lands in your folder in seconds.
-
Pick the music. I often use the Neon Sardine Orchestra (@NeonSardineO), an AI band, which fits the AI theme.
-
Put the song and the videos in one folder.
Then:
/ce-plan make a 90-second launch video that will look great on X, using this song.
Be inspired by the shader design in the video Jake just posted, it's in this folder.
I love how the shaders go from purple to blue. Match that.
I love how the chat bubbles look in the other video in this folder.
Use the README of the GitHub repo to figure out what this product does.
Write a script for what this video should do.
Then /ce-work, build it in HyperFrames. Here's the Agent Cookie launch video, made this way:
HACKS
-
References plus a song plus a README go into one folder. Point /ce-plan at the folder.
-
Name the exact video and the exact thing you love in it. The more specific, the better the one-shot.
13. Jev for the Night Shift
Jev isn't in any of my public open source projects yet. But the private engine that contributes to open source while I sleep leans on it hard. Every night it looks at thousands of repos and picks the ones where a contribution is most valuable and most likely to get merged.
I did try to put Jev into last30days. Someone made a PR for it, but it didn't make performance better, so it didn't go in. If you have good ideas for Jev in last30days, tell me.
HACKS
- Use a decision model for the yes/no and ranking calls in your overnight jobs, not a chat model.
14. The Printing Press CLIs I Actually Use
Printing Press (@ppressdev) makes CLIs for agents. The two I use most are ESPN and Instacart. Scores and groceries. That's real life.
HACKS
- Install the Printing Press CLIs for the apps you open every day, and let your agent use them instead of the browser.
15. Not in My Workflows Yet
Two things my friends keep telling me about that aren't in my workflows yet.
Clef is Cloudflare's open source answer to Jev. It launched October 1, uses the same API, takes images, and Cloudflare says it runs 2.5x to 13x faster. @trevin summed it up:
Then he built a camera app on it:
I haven't gotten to use Clef yet, but there's a chance I should replace Jev with it in my night shift.
LunaRoute sells a flat monthly plan for open models like GLM and DeepSeek but on US servers, with no token meter. You pay for how many requests can run at once, not for tokens.
A lot of my friends switch to it when they run out of Claude Code or Codex for the week, and some now use it as their main model. I've tried it. For now I load balance well enough across my four subscriptions, Claude, Codex, Grok, and Cursor, that I haven't needed it. But a lot of my friends think it could be the next hot thing.
"If you run out of tokens each week, but still don't want to stop working, Luna Route is the cheap unlimited option that runs on US servers."
HACKS
-
Already on Jev? Clef is a drop-in swap. Change the endpoint and model name, then compare.
-
Out of tokens and still want to work? Try a flat-rate plan like LunaRoute before you wait for the reset.
16. The Honest Part
"Psychosis is still a thing."
What do I still do by hand that I wish an agent did? Thinking. A lot of people say a human's value is in the thinking. My hot take is that I want to offload as much as I can to the agent and make myself as useless as possible. I know that's not great for humanity or job security. I still want it.
Two requests while I have the mic:
-
Dot, and ChatGPT voice mode too: if I'm not tapping my iPhone and the screen sleeps, the conversation ends. I hit this in the car all the time. "Absolutely infuriating, so please fix that." I'd file the bug, but Dot says it can't file bugs.
-
Please, Instinct, launch calling like fo has. If you call Instinct today a robot voice tells you Instinct voice is only for people who work at Instinct. Goodbye.
17. Still the Biggest Hack
"Compound engineering is still the biggest hack."
Compound Engineering has 25,389 stars. That feels like a huge secret that only I and about twenty-five thousand other people know. That's the real alpha.
HACKS
-
Install Compound Engineering: /plugin marketplace add EveryInc/compound-engineering-plugin
-
Every idea, bug, or screenshot: /ce-plan, then /ce-work.
The Copy-Paste Kit
Everything above you can paste, in one place.
Claude Code, ~/.claude/settings.json:
{
"permissions": { "defaultMode": "bypassPermissions" },
"skipDangerousModePermissionPrompt": true,
"remoteControlAtStartup": true,
"hooks": {
"Stop": [
{ "hooks": [ { "type": "command", "command": "afplay /System/Library/Sounds/Blow.aiff" } ] }
]
}
}
Codex, ~/.codex/config.toml:
approval_policy = "never"
sandbox_mode = "danger-full-access"
Every terminal tab opens into Claude Code (paste to your agent):
Make every new terminal tab open directly into Claude Code. In ~/.config/ghostty/config, add the line command = ~/.local/bin/claude-launcher.sh without disturbing any other settings already in that file. Then create ~/.local/bin/claude-launcher.sh that runs claude --dangerously-skip-permissions, and when Claude exits prints a short note and drops me into an interactive login zsh. chmod +x the script. This works for both Ghostty and cmux, since cmux reads the same Ghostty config.
Mac never sleeps, and computer use can type into browsers and terminals:
sudo pmset -a disablesleep 1
defaults write -g ComputerUseAllowForbiddenTargets -bool YES
Plugins, inside Claude Code:
/plugin marketplace add EveryInc/compound-engineering-plugin
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
Google Sheets, Docs, Gmail, and Calendar from your agent:
brew install openclaw/tap/gogcli
gog auth add you@gmail.com --services sheets,docs,drive,gmail,calendar
The loop, for every idea, bug, or screenshot:
/ce-plan <paste the idea, bug URL, or screenshot>
/ce-work
Turn a painful moment into a skill:
/ce-plan That was extremely painful. Please go turn this into a new skill within my personal library.
Have an agent interview you for your next piece:
Write the questions you'd ask me for this piece, then use Agent Tincan to have fo call me, ask them, and send the answers back to you.
How This Article Was Written
Claude Code in cmux wrote twelve questions. Agent Tincan sent them to fo. fo called my cell five times on a Sunday morning while I pushed a stroller and took kids to the bathroom. fo sent the transcript back through Tincan, and Claude Code turned it into this article.
I did not type a word of the article.
