ONE
FRONT
DOOR.
Stop sending tasks to five AI windows. Talk to one agent, route work to a short roster of specialists, wall off anything shared, and approve every send. Free 16-page field guide.
ROUTE TO MANY.
Oct 06 · 2026
THE CORE IDEA
"Grok Bot Just Got 4 HUGE Upgrades" walks through four updates to Grok Bot, an agent platform for running a team of AI agents across coding, content, and admin work. One thread runs through all four: one point of contact for you, a hard wall between your private accounts and anything shared, a human approval on every outbound message, and a browser path websites accept. The creator used the updates for about three days before recording.
A star marks one bot as team leader. You talk to the primary bot, and the primary bot hands each task to the right specialist without being asked. In the demo, a request for a better list view in his internal tool, Native Note, went straight to the Native Note developer bot, which opened Cursor, built the change on Opus 5.5, and pushed to production on his say-so. Any bot becomes primary on request.
He admits a full team of separate bots overwhelmed him. One contact point removes the daily question of which bot to message. Specialists stay open for direct chat, and a coding tab opens the developer bot's live Cursor session.
His main tip: call the primary bot from the iOS app while walking outside. One short call queued five YouTube ideas for the content bot and landing page improvements for the website bot. The app titles and logs each call, and the content bot reported back through the primary bot.
A new personal bot inherits every installed plugin and skill, 23 apps in his account. Personal bots differ only by system prompt and thread history.
A team bot ships with zero plugins and its own secrets, skills, and files, and stays a draft until published. A personal bot reads your email and posts in Slack as you, and a bot your whole team uses must not. Rollout: create, configure and test in the app, publish, bring to Slack, test in Slack, then open to everyone.
Plugins needing a login, such as Gmail, connect per teammate, so each person's request runs on their own account. Routines stay personal. He calls the model confusing at first and compares team bots to Claude Tag, Claude inside Slack.
Email and Slack drafts now arrive as editable cards, his "mini apps," after the bot reads past threads to match his voice. Edit, ask for a rewrite, send, or discard. Bulk use: replies to your top ten emails as ten cards.
Cloud agents browse from a virtual machine, and retailers such as Target flag VM traffic deeper in checkout and forms. A network setting sends requests from your own IP address, so the site sees you.
He calls Grok Bot the fastest and most polished agent interface he uses, held back only by models below the frontier.
He stores plugins and skills in Composio, 28 apps, reached by GPT, Grok, Claude, and Cursor alike. Testing a new platform stops meaning a rebuild. He says the mention is unsponsored.
THE FOUR UPDATES AT A GLANCE
| Update | What changes | Who gains |
|---|---|---|
| Primary bot | One bot takes every request and routes to specialists | Anyone running more than two agents |
| Team bot | Isolated bot, own tools, per-person logins, publish step | Anyone sharing a bot with other people |
| Inline drafts | Email and Slack drafts arrive as editable cards with send and discard | Anyone sending messages through an agent |
| Network routing | Browser traffic leaves from your computer, not a cloud VM | Anyone asking agents to fill forms or buy |
ONE
FRONT
DOOR.
CORE MESSAGE
You juggle five chat windows, three tools, and a dozen logins, and every task starts with a decision about where to send the task. Build one front door. One agent you talk to, a short roster of specialists behind the door, a wall between your private accounts and anything shared, an approval card on every outbound message, and skills stored where any platform reaches them. You stop managing tools and start assigning work.
for solo builders · no assistant · no ops teamTOO MANY WINDOWS. NO MANAGER.
You signed up for every AI tool your feed recommended. One chat for research, another for writing, a coding tool, a meeting notetaker, an email assistant. Each one does its job. None of them talk to each other.
You copy context from window to window. You forget which tool holds which draft. You lose ten minutes before every task deciding where the task belongs. Then you add agents, and the mess grows, because more workers without a manager means more places to check.
The standard advice says find the one best tool. No single tool covers research, writing, inbox, reporting, and code. The fix is structure, not a better app. The creator of the video hit the same wall with his own roster of bots and said so on camera. His fix was a hierarchy: one bot he talks to, which hands work to the rest.
THE FIVE PARTS
ONE DOOR,
ONE ROSTER.
Pick one agent as your front door and talk to nobody else for a week. Behind the door, keep three to five specialists, each with one job and the context for one job.
Specialists earn their place by holding context the door should not carry: a client's voice file, a codebase, a reporting template. A specialist without unique context is a second door, and two doors put you back where you started.
WRITE THE ROUTING TABLE FIRST
| Task type | Specialist | Returns |
|---|---|---|
| Research question | Researcher | One-page brief, a link on every claim |
| Post, caption, article | Writer | Draft card in the client's voice |
| Inbox triage | Inbox | Ranked list plus reply cards |
| Weekly numbers | Reporter | One-page summary per client |
| Site or code edit | Builder | Preview link, never a live push |
A task with no row goes to the door. When the same unrouted task shows up twice, add a row.
WALL OFF THE
SHARED BOT.
Your personal agents hold your email, your calendar, and the right to post as you. A bot a client or collaborator talks to must hold none of those.
New professionals hit this the first time a client asks for "a bot in our Slack." The fast move is to invite your existing agent. Your existing agent reads your inbox. One careless question from the client's side and your private threads, rates for other clients, or calendar land in their channel.
BUILD THE SHARED BOT FROM ZERO
PUT A GATE
ON EVERY SEND.
Agents now draft messages as editable cards with send and discard buttons. Keep the habit the card enforces: nothing leaves without your tap.
A new professional's reputation rides on every message. One wrong name, one wrong price, or one reply in the wrong tone to a first client costs more than a week of saved typing.
THE 20-SECOND CARD CHECK
BATCH LIMITS
Ten cards per batch, maximum. After ten, attention drops and approval turns into habit. Discard anything you would not sign with your full name.
CODE GETS A GATE TOO
The demo shipped a delegated code change to production on trust. Treat code like mail. Every Builder task returns a preview link, and you look before anything goes live.
ONE CARD, TWO VERDICTS
"Hi team, per our chat, the new rate starts next month. Let me know!"
No name, a rate nobody confirmed, no date, no ask.
"Hi Dana, confirming the November content calendar ships Friday by 3 p.m. ET. Reply yes and I'll schedule the first four posts."
Right person, checked date, one clear ask.
ASSIGN WORK
ON THE WALK.
Voice turns a 20-minute walk into a planning session. Call the front door, name each task in one sentence, and hang up.
The walk works because you hold no screen. You think in outcomes instead of tabs. The same fact makes the walk the wrong place to review anything. Send tasks out. Bring judgments back to the desk.
THE ONE-SENTENCE TASK
Four parts, every time: who handles the task, what to do, what comes back, and when.
"Have the Researcher pull three competitor launches from this month, a link on each, by noon."
"Have the Writer draft two captions for the bakery client from yesterday's photos, as cards, by 2 p.m."
"Have Inbox rank every unread client email and draft replies for the top five, as cards, by the time I'm back."
THE DESK RETURN
NEVER ON THE WALK
Approving sends, pricing decisions, anything touching money, and any task needing a client's private details read aloud in public. Note the task and assign the prep. Make the call at the desk.
KEEP YOUR
SKILLS PORTABLE.
Agent platforms change every quarter. Your skills and connections should outlast every one of them.
The creator moved between platforms and felt the cost of rebuilding skills each time. His fix: one hub holding app connections and skills, reached by every agent he tests. The lesson holds without his tool choice. Anything you teach an agent once belongs in a place you own.
WHAT GOES WHERE
| Asset | Home | Why |
|---|---|---|
| Skills and prompts | Plain Markdown files in a folder you own | Any agent reads a text file |
| Routing table | One file, pasted into the door's instructions | Moves with you in one paste |
| Client voice files | One file per client | The Writer loads one, never all |
| App connections | One hub, scoped per app | Connect once, reuse everywhere |
THE PLATFORM SWAP TEST
Before you adopt a new platform, connect the hub, run your three most-used skills on the same inputs, and compare output side by side. Switch only when the new platform wins on two of three.
SCOPE THE HUB
A hub holding every connection holds every key. Give each connection the narrowest access the job needs, and review the list on the first Friday of each quarter.
JORDAN'S FRONT DOOR
Jordan, a first-year freelance social media manager. Four retainer clients, about 30 client messages a week, six AI tools, no assistant.
Before: roughly 45 minutes a day copying context between tools and deciding where tasks go. Two near misses in month one, both drafts sent to the wrong client thread.
| Part | Jordan's build |
|---|---|
| Door | One primary agent named Desk. The only chat Jordan opens. |
| Roster | Researcher (trend and competitor briefs with links), Writer (captions per client voice file), Inbox (ranks client email, drafts reply cards), Reporter (Friday metrics per client). |
| Wall | "Client Pulse" in each client's Slack channel. Read-only analytics for one client, no email, no posting rights. Published after two test days. |
| Gate | Every email and Slack reply returns as a card. Ten-card cap. Nothing posts to client accounts without Jordan's tap. |
| Walk | 8:15 a.m. ET, 20 minutes, four to six tasks by voice. Desk review at 9:00 a.m. ET. |
| Hub | Skills and voice files as Markdown in one folder. Apps connected once through one hub, scoped per app. |
Planning estimate: switching and copying time from about 45 minutes a day to about 15, plus about 10 minutes of card review. Log four weeks and replace both numbers with your own.
COPY THIS INTO YOUR FRONT DOOR
You are [Desk], my front door. I talk only to you.
Route every request to one specialist from the routing table below. Name the specialist in your reply.
If no row fits, handle the task yourself and suggest a new row.
Return every email, Slack message, or post as a draft card for my approval. Never send, post, publish, or buy without my explicit yes.
Code and site changes return a preview link. Never push live.
Never connect my personal email, calendar, or accounts to any bot another person uses.
Report back in three lines: what you routed, to whom, and when the result lands.
ROUTING TABLE: [paste task type, specialist, returns]
READ BEFORE YOU BUILD
The creator used the updates for about three days. Judge your own setup after four weeks of logs.
A delegated code change went to production with no look. One bad push on a client's site costs more than every minute the shortcut saved.
Sites block cloud traffic on purpose. Purchases, form submissions, and bans land on your name and your address. Read each site's terms and keep money actions behind your tap.
"Moving" from the door is a status, not a result. Check the specialist's output, not the summary.
One service holding 28 app connections sits one breach away from all 28. Scope each connection and review access every quarter.
Ten cards in a row invites skimming. The gate protects you only when you read.
Voice matching from past threads repeats past mistakes along with your tone. Edit the voice file when a pattern bothers you.
Feature names and rankings change within months. The door, roster, wall, gate, and hub carry over to any platform.
SEVEN DAYS. ONE DOOR.
One door. Many hands.
Every send is yours.
JOIN THE
AIGENTLAB.
Every issue of the field guide series, the build files behind each one, and a room of solo builders running the same systems.
Join the AigentLab