{"id":3077,"date":"2026-09-15T00:00:00","date_gmt":"2026-09-14T23:00:00","guid":{"rendered":"https:\/\/contentlabstudy.com\/soft\/grokbot-ai-teammates-setup-skills-multi-bot-workflows\/"},"modified":"2026-09-15T00:00:00","modified_gmt":"2026-09-14T23:00:00","slug":"grokbot-ai-teammates-setup-skills-multi-bot-workflows","status":"publish","type":"post","link":"https:\/\/contentlabstudy.com\/soft\/grokbot-ai-teammates-setup-skills-multi-bot-workflows\/","title":{"rendered":"Grokbot AI Teammates: Setup, Skills, Multi-Bot Workflows"},"content":{"rendered":"\n<video controls playsinline preload=\"metadata\"\n       style=\"width:100%;max-width:960px;height:auto\"\n       poster=\"https:\/\/contentlabstudy.com\/soft\/wp-content\/uploads\/2026\/09\/image_2026-09-15-02-Grok-Bot-101.jpg\">\n  <source src=\"https:\/\/contentlabstudy.s3.us-east-1.amazonaws.com\/soft\/2026-09-15_02_Grok_Bot_101.mp4?X-Amz-Algorithm=AWS4-HMAC-SHA256&amp;X-Amz-Credential=AKIAVTCJZPR4S4LPREPL%2F20260921%2Fus-east-1%2Fs3%2Faws4_request&amp;X-Amz-Date=20260921T102035Z&amp;X-Amz-Expires=21600&amp;X-Amz-SignedHeaders=host&amp;response-cache-control=no-cache&amp;X-Amz-Signature=d8ecca3be0050791b0819fe5f7946f1d630b5b2b485183d3293acadc403ddd37\" type=\"video\/mp4\" \/>\n<\/video>\n\n<!--\nTitle: Grokbot AI Teammates: Setup, Skills, and Multi-Bot Workflows\nSlug: grokbot-ai-teammates-setup-workflow\nMeta description: Set up Grokbot as cloud AI teammates that own outcomes, drive a computer, learn skills, and finish forms, slides, and email with human approvals.\nExcerpt: Grokbot is built as a cloud teammate, not a chat window: you create bots for job responsibilities, give them a computer, and let them finish the work. This operating manual covers role-based bots, computer use, skills, approval rules, and a multi-bot loop from forms to slides to email. Start with one annoying task, then add memory, routines, and a manager bot.\nTags: Grokbot, AI teammates, knowledge work, agent workflows, enterprise AI, computer use\nEditor notes: Uncertain names: Email Ethan is also called \u201cEmil Ethan\u201d in the demo (used Email Ethan); Slide Sonia is once called \u201cSlide Guru\u201d; marketplace names appear as \u201cClairevo\u201d then \u201cClaire Lawrence\u201d (spelling uncertain); \u201cJason Demore\u201d as spoken; \u201cOne Password\u201d as spoken (likely 1Password); bot messages say \u201cRita\u201d for Amrita. Speakers mix \u201cSpaceX AI\u201d and \u201cCursor\u201d as their internal context\u2014left as stated. Measurement section thinned: no time-saved, cost, or lead numbers. No verbatim system prompts beyond the Google Form request; other prompt blocks are reconstructed from spoken instructions. Suggested video embed: after Key takeaways. Verify: voice back-and-forth timing, One Password integration, public marketplace roster, Starbase\/Starship challenge, multiplayer human+bot chat \u201ccoming very soon,\u201d x.ai\/galaxy session list.\n-->\n<h2>Key takeaways<\/h2>\n<ul>\n<li>Create a Grokbot for a job responsibility you will return to, not a one-off chat for each task. Roman\u2019s rule: teammates sit at the center, not tasks.<\/li>\n<li>Give the bot an outcome, not a pile of steps. Amrita\u2019s operating habit is to state what you want produced and let the bot work backwards.<\/li>\n<li>When a tool has no usable MCP or API, Grokbot still works: it drives its own cloud Linux VM the way a person would, including Google Forms, Qualtrics, older software, video, and audio.<\/li>\n<li>Write approval rules before the bot talks to the outside world. Default auto-review already flags risky actions; you can require permission for email, Slack to customers, or even mundane computer steps.<\/li>\n<li>Teach by doing. Control the bot\u2019s computer, demonstrate a motion such as a slide animation, stop the recording, and Grokbot stores it as a reusable skill.<\/li>\n<li>Long-lived personas beat project-scoped chats. Memory, descriptions, and routines let Email Ethan, Slide Sonia, or a brand bot work across many decks and inboxes.<\/li>\n<li>Orchestrate specialists instead of stuffing one bot. Message bots to each other, put them in a group chat for visibility, or add a manager bot that solicits blockers on a schedule.<\/li>\n<li>Share a bot as a template when you need consistency (brand, tone, fonts). Duplicate a bot when you need the same persona with a clean memory.<\/li>\n<\/ul>\n<h2>Who this is for (and who should skip it)<\/h2>\n<p>This method is for a founder, operator, marketer, GTM lead, or ops person who already has (or can get) Grokbot and wants finished work: forms, sheets, slides, and drafts that can run while a laptop is closed. The live path assumes Google Forms, Slides, Sheets, and email. Optional pieces the speakers used: MCPs\/plugins, Qualtrics, Slack, a VPN, MongoDB credentials, and One Password for secrets. Comfort with granting a bot a computer, watching it click, and writing a few rules is enough. You do not need to be an engineer, though engineering and admin sessions exist separately at x.ai\/galaxy.<\/p>\n<p>Skip this if you only want a research chatbot, if you cannot give a bot credentials or computer access, or if you need a tool-agnostic playbook. Grokbot is the stack. Roman\u2019s bet is specifically that knowledge work should look more like how developers already steer agents: you own direction and outcomes; the bots do the mundane path from A to B.<\/p>\n<h2>Treat Grokbot as a teammate, not a chat tab<\/h2>\n<p>Roman frames three phases of workplace AI. First, chat: you ask, it answers, research gets faster. Second, copilot: you delegate a discrete task and iterate, for example on a slide deck. Third, the moment Grokbot is built for: AI that behaves like a colleague. You are not outsourcing a checklist. You are assigning a lane of the company, with responsibilities, memory, and trust that the work comes back finished.<\/p>\n<p>That product bet shows up in three design choices.<\/p>\n<ul>\n<li><strong>Chat-like teammates.<\/strong> The left rail is a roster of bots with roles (sales outbound, inbox manager, Data Dan), not a thread per task. You come back to the same bot. It is supposed to remember, weeks later, that you like slides formatted a certain way.<\/li>\n<li><strong>A computer plus integrations.<\/strong> MCPs and APIs help, but they often stop at about 90%. Grokbot also has its own computer and can click through software a human can use, including tools with no modern API.<\/li>\n<li><strong>Cloud, always on.<\/strong> Work does not live on your laptop. If you are asleep, the lid is closed, or you kick something off from your phone, the bot still runs.<\/li>\n<\/ul>\n<blockquote>You actually create a bot for a unit of work and a kind of job responsibility.<\/blockquote>\n<p>Amrita\u2019s demo is the same idea in motion: three specialists on one project. Data Dan collects user data. Slide Sonia turns it into slides and charts. Email Ethan packages the analysis for an external stakeholder. You participate as the manager, not as the person building every form field by hand.<\/p>\n<table>\n<thead>\n<tr>\n<th>Use case<\/th>\n<th>Bot \/ tool<\/th>\n<th>Input needed<\/th>\n<th>Output<\/th>\n<th>Owner<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Collect survey or research data<\/td>\n<td>Data Dan on Google Forms (Qualtrics works the same way)<\/td>\n<td>Questions, multiple-choice options, sharing settings<\/td>\n<td>Live form, QR code, responses in a Google Sheet<\/td>\n<td>Ops, user research, event ops<\/td>\n<\/tr>\n<tr>\n<td>Turn data into a deck<\/td>\n<td>Slide Sonia on Google Slides<\/td>\n<td>Sheet or form data, fonts\/colors, taught skills (e.g. animation)<\/td>\n<td>Edited deck, chart, screenshot of the finished slide<\/td>\n<td>Marketing, GTM, anyone presenting<\/td>\n<\/tr>\n<tr>\n<td>Brief a stakeholder<\/td>\n<td>Email Ethan with Google \/ email<\/td>\n<td>Context from other bots, sample emails for tone, recipient<\/td>\n<td>Draft in a modal; send only after your rule allows it<\/td>\n<td>Founder, GTM, customer-facing lead<\/td>\n<\/tr>\n<tr>\n<td>Keep the team unblocked<\/td>\n<td>Manager \/ chief-of-staff bot<\/td>\n<td>Named reports, a cadence, what \u201cblocked\u201d means<\/td>\n<td>Status, blockers, pings to you or sibling bots<\/td>\n<td>Operator overseeing the roster<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>Create a Grokbot for a job and give it a first outcome<\/h2>\n<p>What you are trying to accomplish: stand up one long-lived bot that owns a lane, then prove it can finish a simple piece of work without you building the artifact yourself.<\/p>\n<p>Prepare a name that sounds like a role. In the demo, \u201cData Dan\u201d is enough for Grokbot to guess a purpose. A name like \u201cworkflow builder\u201d likewise seeds the bot with an idea of the job. You will also see prompts for \u201cwhat do you mainly want me for.\u201d You can pick one, or ignore them and go straight to the outcome.<\/p>\n<h3>Create the bot and assign the first job<\/h3>\n<ol>\n<li>Create a new bot and name it for the responsibility (Data Dan, not \u201cform for Thursday\u201d).<\/li>\n<li>State the outcome in one request, typed or via voice. Voice mode in the session was speech-to-text; Amrita said a back-and-forth voice mode was deployed internally and expected out around the following week.<\/li>\n<li>Let the bot start immediately. Do not open a new chat for the next subtask. Stay in this bot.<\/li>\n<\/ol>\n<p>The Google Form request used in the room:<\/p>\n<pre><code>Create a Google form for me that asks two questions. The first is how many cups of coffee do you drink a day? And the second is what are your favorite coffee shops in San Francisco? And for the coffee shop options, maybe put Ritual, Fields, Sightglass, and Blue Bottle.<\/code><\/pre>\n<p>Good looks like: the bot begins working without a long planning essay, then uses its computer to build the actual form. Failure looks like a private link nobody can open, or a form that asks for data you did not intend to collect. When the QR code in the demo was unscannable, Amrita did not restart from scratch. She told Data Dan the outcome (\u201cfolks are unable to scan; make sure this link is public\u201d) and let it fix sharing settings.<\/p>\n<h3>Watch the computer, then take over if it stalls<\/h3>\n<p>Open the bot\u2019s computer from the top right. You are looking at an isolated Linux VM, not your laptop. It can log into Google Forms, Slides, Docs, and Sheets. One bot cannot see another bot\u2019s machine, which is why Data Dan can build a form while Slide Sonia works a deck at the same time.<\/p>\n<p>If a step is slow or wrong, do it yourself on that same computer, then teach the bot the motion so the next run is shorter. Amrita uses this for long-running UI work and for anything without an MCP: Google Forms, Qualtrics, and similar. She also uses the VM on airplanes because its network is fast even when hers is not.<\/p>\n<p>Computer-use speed, she noted, still depends on the model and on ongoing computer-use work. Treat delays as a reason to watch, approve, or demonstrate, not as a reason to abandon the bot.<\/p>\n<h2>Connect MCPs and set Grokbot approval rules<\/h2>\n<p>What you are trying to accomplish: give the bot the shortest path for tools that already have plugins, and a hard gate for anything customer-facing or irreversible.<\/p>\n<p>In the bot, open the MCP \/ plugin list and connect what the job needs, or ask Grokbot to set a connection up. Use MCPs when they exist. Use the computer when they do not. You do not have to pick only one.<\/p>\n<p>Approvals show up as \u201cthe bot wants to run a task.\u201d Allow once while you are learning. Then encode policy in settings so you are not tapping Allow all day.<\/p>\n<h3>Write rules that match risk, not convenience<\/h3>\n<p>Amrita\u2019s examples:<\/p>\n<ul>\n<li>Do not reply to email unless the bot explicitly asks permission first.<\/li>\n<li>Allow slide creation automatically, without asking.<\/li>\n<li>Block a send that would go to an external customer until you approve.<\/li>\n<\/ul>\n<p>Under the hood there is also an auto-review classifier that scores how risky an action is. In the demo it was conservative: creating a coffee form still triggered checks so the form would not collect PHI or PII, and connecting to outside services required permission. You can go stricter than auto-review. You can require approval even for a new folder on the desktop. You can also loosen a pattern after you have seen good work.<\/p>\n<p>A question from the room: if the bot stops you constantly, it never works on its own. Amrita\u2019s answer is memory. If it asks whether copy is good enough to send, you can reply that the copy is fine and that going forward it should send without asking about copy. That is how trust accumulates. The heavy approval loop in the live demo, she said, was not how her bots usually behave.<\/p>\n<p>Good looks like: slides move without you; email drafts appear in a modal; sends wait. Failure looks like a bot blocked by your own auto-review rule while you wonder why nothing went out. Read the block, allow once if you mean it, then decide whether the rule should stay.<\/p>\n<h2>Teach skills, write the job description, and add routines<\/h2>\n<p>What you are trying to accomplish: make the second and tenth run look like the same colleague, not a new intern.<\/p>\n<h3>Teach a task from the bot\u2019s computer<\/h3>\n<p>On Slide Sonia, use <strong>Teach a task<\/strong>. You control the bot\u2019s computer. In the demo that meant Insert \u2192 animation \u2192 fly-in, then play it to confirm timing. Stop the task. Grokbot turns the recording into a skill you can call later (\u201cadd that animation to this text or image\u201d).<\/p>\n<p>What is happening: the bot watches what you showed, breaks it into the clicks (select object, pick type, optional preview), and saves a named skill. You are not writing a tutorial. You are demonstrating once on its machine.<\/p>\n<h3>Put non-negotiables in the bot description<\/h3>\n<p>Amrita keeps Slide Sonia consistent by writing explicit instructions: how to edit an existing deck, which font and colors to use, and a requirement to screenshot the finished slide and send it back. That screenshot is the QA loop. You review the image instead of hunting through the file.<\/p>\n<p>Use descriptions for swim lanes too. People she has talked to run a chief-of-staff bot whose description says it should always delegate to the team and not do the work itself.<\/p>\n<h3>Routines and memory<\/h3>\n<p>Add a routine on Sonia such as: every morning at 9:00 a.m., summarize changes to this shared deck and who made them. Routines can span bots: Slide Sonia and Email Ethan email your boss a weekly summary of deck changes. A manager bot can, every two hours, solicit updates and blockers from Email Ethan, Slide Sonia, and Data Dan.<\/p>\n<p>Anytime you steer (\u201cnever use my first and last name; just my first name\u201d), treat that as memory, not a one-time correction. Amrita said that guidance is stored for later choices. Memory itself lives in S3: long-lived and persistent, and you can modify it. If you spun up bots for a demo and then deleted them, tell the chief of staff to stop delegating to them. A bot will not automatically know a sibling is gone. Duplicate a bot when you want the same persona with empty memory, for example a fresh Email Ethan who still knows how to write but has none of last quarter\u2019s threads.<\/p>\n<p>When Grokbot refuses or asks a question, take it as a signal to teach or guide, not as a dead end.<\/p>\n<h2>Orchestrate multiple Grokbot AI teammates on one project<\/h2>\n<p>What you are trying to accomplish: finished work across tools you do not want to click through, while you stay on orchestration.<\/p>\n<p>You can divide bots by expertise (slides, email, data) or by project. Amrita\u2019s preference in the demo is experts with memory who can be reused for a year. Email Ethan can learn from every email you have sent. Slide Sonia\u2019s job is slides generally, including a new brand or theme you teach later. Duplicate Email Ethan if stakeholders need different tones: one for execs and CFOs, one for end users, one for your internal team.<\/p>\n<h3>Bot-to-bot messages, then a group chat for visibility<\/h3>\n<p>Two patterns:<\/p>\n<ul>\n<li><strong>Direct handoff.<\/strong> Tell Email Ethan to message Slide Sonia for details on a named deck, or to message Data Dan for the coffee numbers. In the demo, Ethan was smart enough to ask Dan for coffee data and Sonia for the Galaxy 101 deck. You will see the inter-bot thread: \u201cRita asked me to do this,\u201d \u201cI don\u2019t have this yet, still on the spreadsheet,\u201d \u201cI\u2019ll use sample data until then.\u201d<\/li>\n<li><strong>Group chat.<\/strong> Put Email Ethan, Data Dan, and Slide Sonia in one chat when you want to watch the protocol without clicking into each 1:1. Someone asked how that differs from bots messaging each other. Amrita\u2019s answer: visibility. Delegation still happens either way. In a group you can also say, in one place, \u201ccreate a chart on the last slide from this sheet; when you\u2019re done, Email Ethan sends Jason the coffee data.\u201d The bots respect dependencies: Ethan holds the email until Sonia\u2019s screenshot lands. She flagged the same pattern as useful in engineering to avoid conflicting changes.<\/li>\n<\/ul>\n<p>You can still edit a shared Google deck yourself on later slides while Sonia works earlier ones. Isolated VMs mean the bots are not fighting over one desktop; the file is the collaboration surface.<\/p>\n<h3>Add a manager when you do not want to be the pager<\/h3>\n<p>Create a bot named Manager. Tell it who it manages and that its job is updates and blockers. If Data Dan cannot produce a shareable form, the manager should surface that and ping you or Sonia. This is the human workflow Amrita is copying: reports escalate, the manager unblocks.<\/p>\n<p>Good looks like: you are not in the nitty-gritty; a draft, a slide, and a form appear; agents wait on each other instead of colliding. Failure looks like you asking \u201cwhat\u2019s going on?\u201d because nothing is in a group chat, no manager is polling, and each bot is a private silo.<\/p>\n<h3>Drafts, QR codes, and sharing with the team<\/h3>\n<p>Email drafts open in a modal: change the recipient, delete, or ask for a rephrase without living in your inbox. For distribution, Amrita prefers QR codes Grokbot generates over mailing raw form links.<\/p>\n<p>Share a bot as a template from the top of the bot. That is the multiplayer path: one brand-design bot with the same colors, fonts, and typography, published so the design team stays consistent, then customized per person. Updates flow to people using the template. The marketplace also has public bots (the session showed published examples from Cursor, including Lauren, Lenny, and Eric). Use internal publishing for company consistency; use the marketplace when a public persona is enough to start.<\/p>\n<p>Categorize bots into sections (engineering roster vs session roster vs unassigned) so the left rail stays a team chart, not a junk drawer.<\/p>\n<h2>Prompts, bot instructions, and workflows<\/h2>\n<p>Use these as copy-ready starting points. They follow what was said in the session, not a rewritten methodology.<\/p>\n<p>Slide Sonia description, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>You edit existing slide decks. Use the specified font and colors. When you finish a slide, take a screenshot of it and send it back to me. Stay consistent with the instructions I have already given you.<\/code><\/pre>\n<p>Chief-of-staff \/ always-delegate instruction, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>You are chief of staff. Always delegate to your team. Do not do the underlying work yourself. Your subordinates have the right context; route work to them and bring back finished results.<\/code><\/pre>\n<p>Manager bot, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Your job is to get updates from Email Ethan, Slide Sonia, and Data Dan on what they are working on. Every two hours, solicit updates from your team and see if there are any blockers. If someone is blocked, surface it and ping me (and the relevant sibling bot) so we can unblock.<\/code><\/pre>\n<p>Daily deck-change routine, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Every morning at 9:00 a.m., summarize any changes that were made to this slide deck and tell me who made those changes.<\/code><\/pre>\n<p>Weekly cross-bot routine, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Slide Sonia and Email Ethan: once a week, work together and email my boss a summary of all changes made to this slide deck.<\/code><\/pre>\n<p>Agent-to-agent steer used in the demo:<\/p>\n<pre><code>Message Slide Sonia and ask her for any details about the Grokbot Galaxy 101 slide deck that she made. I want to send an email to my coworker about what we showed today.<\/code><\/pre>\n<p>Group-chat orchestration used in the demo, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Create a new slide in Slide Sonia that has data from the coffee spreadsheet. Create a chart on the last slide of the deck. When you are done, Email Ethan should send an email to Jason with details about the coffee data.<\/code><\/pre>\n<p>Memory steer:<\/p>\n<pre><code>I want you to make sure that you never use my first and last name whenever referring to me. Just use my first name.<\/code><\/pre>\n<p>Debug steer used when the QR code failed:<\/p>\n<pre><code>Looks like folks are unable to scan. Let's make sure this link is public.<\/code><\/pre>\n<p>Email permission rule, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Do not reply to emails for me unless you explicitly ask for my permission first. You may create slides automatically without asking for permission.<\/code><\/pre>\n<p>Tone-training for Email Ethan, reconstructed from the speaker&#8217;s description:<\/p>\n<pre><code>Look at real existing emails I have written and consolidate that into a skill and tone you will use for further emails. Draft an email to [contact] about the data I gathered on coffee habits.<\/code><\/pre>\n<h2>Measurement and business impact<\/h2>\n<p>The speakers did not give time-saved percentages, cost figures, or lead counts. Judge the system the way they did in the room.<\/p>\n<ul>\n<li><strong>Finished vs almost.<\/strong> Roman\u2019s bar is end-to-end work you can trust, not a 90% draft that still needs you to click the last mile.<\/li>\n<li><strong>It runs with the lid closed.<\/strong> If the job dies when your laptop sleeps, you are still in the old local-automation world.<\/li>\n<li><strong>You manage; they execute.<\/strong> Amrita\u2019s checkpoint is whether she is watching orchestration or rebuilding form fields, animations, and email copy by hand.<\/li>\n<li><strong>QA artifact.<\/strong> A screenshot of the slide, a draft modal, a public QR code, and a sheet of responses are the proof, not a status paragraph.<\/li>\n<li><strong>Fewer interrupts over time.<\/strong> If auto-review still stops every form field after you have approved the pattern, memory and rules are not doing their job yet.<\/li>\n<\/ul>\n<p>The motivating workloads were ordinary and heavy: nobody wants to build Google Forms; an event planner collecting dietary restrictions for a large onsite; a user researcher gathering customer data; outbound email; engineering chores such as cleaning stale feature flags; field engineers catching forgotten support tickets. The business outcome Amrita named is time back from the most annoying part of the day, spent on work you actually want.<\/p>\n<blockquote>You give it outcomes, it delivers.<\/blockquote>\n<h2>Pitfalls and guardrails<\/h2>\n<ul>\n<li><strong>Task chats instead of roles.<\/strong> If you open a new bot per request, you throw away memory, skills, and the teammate model.<\/li>\n<li><strong>Vague debugging.<\/strong> \u201cThis isn\u2019t working\u201d does not teach. State the outcome, what failed, and confirm when the bot troubleshot the right way. Amrita\u2019s field-engineering habit is more detail plus validation.<\/li>\n<li><strong>External actions without rules.<\/strong> Email, payments, and talking to an outside company are the cases the audience worried about. Use auto-review plus explicit rules. Keep a human in the loop for sends until memory has earned a narrower exception.<\/li>\n<li><strong>PHI and PII in \u201charmless\u201d forms.<\/strong> The auto-reviewer in the demo cared about this even on a coffee survey. You should too.<\/li>\n<li><strong>Secrets on the VM.<\/strong> For One Password and similar logins, Grokbot is supposed to pop a modal, store credentials securely inside Grokbot, and not expose them on the VM or to the SpaceX side. Do not paste API secrets into the chat as ordinary text if the modal path exists.<\/li>\n<li><strong>Internal systems without auth design.<\/strong> The VM can be locked down or open. It can sit on your VPN. It can use MongoDB if you give it credentials. That is a feature and a risk. Grant the least access the job needs.<\/li>\n<li><strong>Stale memory.<\/strong> Deleted bots, old delegations, and one-off demo data will be remembered until you say otherwise. Duplicate when you need a clean slate.<\/li>\n<li><strong>Brand drift.<\/strong> If every marketer has a private slide bot, fonts diverge. Publish one template.<\/li>\n<li><strong>Watching every click.<\/strong> Isolation and group chats exist so you are not babysitting two desktops. If you cannot look away, you have not set rules, screenshots, or a manager bot.<\/li>\n<li><strong>Humans in the Grokbot group chat.<\/strong> Not the current interface. Today the analog is tagging a bot in Slack so multiple people can add requirements in-thread. A multiplayer human-and-bot group chat inside Grokbot was described as coming soon.<\/li>\n<\/ul>\n<h2>7-day implementation plan<\/h2>\n<p><strong>Day 1.<\/strong> Pick the most annoying recurring job on your calendar (forms, outbound email, slide cleanup, forgotten tickets). Create one bot named for that job. Give it a single outcome. Watch the computer. Approve what auto-review flags. Do not build a roster yet.<\/p>\n<p><strong>Days 2\u20133.<\/strong> Connect the MCPs that actually shorten the path. Write two approval rules: one tight (email \/ external), one loose (internal artifacts such as slides). Teach one skill from the bot\u2019s computer. Put fonts, tone, and \u201cscreenshot when done\u201d into the description. Steer memory once on purpose (how it should address you, what it must never send).<\/p>\n<p><strong>Days 4\u20135.<\/strong> Add the second and third specialists those handoffs need. Practice one direct bot-to-bot message and one group chat with a dependency (B waits for A\u2019s screenshot). Run one thin end-to-end loop: collect \u2192 chart or slide \u2192 draft email behind a permission rule. Duplicate an email bot if two audiences need different tones.<\/p>\n<p><strong>Days 6\u20137.<\/strong> Add a 9:00 a.m. or weekly routine. Stand up a manager bot on a two-hour update cadence if you already cannot watch the group chat. Share the best bot as an internal template if someone else should use the same brand or process. Curate memory: tell bots which siblings are gone, and duplicate any persona that has picked up demo junk. Only then consider a public marketplace template.<\/p>\n<h2>Get the annoying work off your calendar<\/h2>\n<p>The point is not a clever coffee survey. It is a roster of cloud colleagues that own a lane, drive a computer when APIs run out, remember how you work, and stop for you when the action is dangerous. Knowledge work gets closer to how strong engineering teams already run agents: you keep outcomes and judgment; they finish the path.<\/p>\n<p>Watch the walkthrough to see computer use, teach-a-task, and agent-to-agent messages, then create one bot for the most annoying task on your calendar.<\/p>\n<h2>FAQ<\/h2>\n<h3>How are Grokbot AI teammates different from chat or a copilot?<\/h3>\n<p>Chat answers questions. A copilot sits next to you on a discrete task. Grokbot is built as a teammate you return to: a bot per job responsibility, with memory, a cloud computer, and an expectation that it owns a lane and finishes work. Roman\u2019s contrast is task paradigm versus teammate paradigm, including the later step of shared AI colleagues inside a company rather than a single-player assistant.<\/p>\n<h3>Do I need MCPs if the bot has its own computer?<\/h3>\n<p>No. MCPs and APIs are useful and already available as plugins, but Amrita\u2019s point is that a lot of real work (Google Forms, Qualtrics, older internal UIs) has no clean MCP. Give the bot the outcome and the computer. Use MCPs when they exist; do not wait on them to start.<\/p>\n<h3>How do Grokbot approval rules work versus auto-review?<\/h3>\n<p>Auto-review is a classifier that decides whether an action is safe enough to run or should stop for you. On top of that, you write rules: always ask before email replies, never ask before slides, ask before Slack to an external customer. You can be more prescriptive than the classifier, including for low-risk computer actions. Loosen a rule only after you have seen the bot do the job well.<\/p>\n<h3>Can Grokbot access internal tools, VPNs, and passwords?<\/h3>\n<p>Yes, if you grant access. The VM can be locked down or open, joined to a VPN, and given credentials for systems such as MongoDB. A One Password integration was described as just released: a modal collects the login, credentials are stored securely in Grokbot, and they are not supposed to sit exposed on the VM. Auth design is on you.<\/p>\n<h3>How does memory work, and can I edit or forget it?<\/h3>\n<p>Memory is persistent (described as S3) and long-lived across projects, which is why a persona can run for a year. You can modify it in conversation, including telling a bot to forget or stop using deleted teammates. Duplicating a bot copies the persona with fresh memory. If you do not curate it, old demo state and dead delegations will keep showing up.<\/p>\n<h3>Should I use a group chat or let bots message each other?<\/h3>\n<p>Both are valid. Direct messages are the handoff. A group chat is for you: one place to see swim lanes, dependencies, and tone. Put specialists in a group when you are orchestrating a multi-step job; keep 1:1 when you only need a single expert.<\/p>\n<h3>Do I need to leave my laptop open for Grokbot to finish?<\/h3>\n<p>No. Cloud VMs are the point. Roman called out local automations that fail at 2:00 a.m. or cannot start from a phone. Amrita\u2019s version: shut the laptop; Data Dan, Slide Sonia, and Email Ethan keep working on forms, slides, and sheets.<\/p>\n<h3>Should I create one Grokbot or many?<\/h3>\n<p>Start with one lane of annoying work. Add specialists when handoffs appear (data vs slides vs email), and add a manager when you cannot watch them. Duplicate for stakeholder-specific tone. Publish one template when the company needs a single brand or process. Many bots is a team design choice, not a scoreboard.<\/p>\n<script>(function(){if(window.__tkS3Vid)return;window.__tkS3Vid=1;document.addEventListener('playing',function(e){var t=e.target;if(t&&t.tagName==='VIDEO'){t.removeAttribute('poster');}},true);}());<\/script>","protected":false},"excerpt":{"rendered":"<p>Set up Grokbot AI teammates that own outcomes, drive a cloud Linux VM, learn skills, and finish the multi-bot workflows with human approvals.<\/p>\n","protected":false},"author":2,"featured_media":3076,"comment_status":"open","ping_status":"0","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[3],"tags":[],"class_list":["post-3077","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-software-development"],"_links":{"self":[{"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/posts\/3077","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/comments?post=3077"}],"version-history":[{"count":0,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/posts\/3077\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/media\/3076"}],"wp:attachment":[{"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/media?parent=3077"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/categories?post=3077"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/contentlabstudy.com\/soft\/wp-json\/wp\/v2\/tags?post=3077"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}