Grokbot for Sales Engineers: Always-On Bots That Finish Work

Sales engineers still lose hours to work that looks small on the calendar: turning a public blog post into a proof-point slide, answering how a booking flow actually behaves in the repo, and checking what a competitor shipped this week. Amrita, a field engineer at Cursor, runs that work through Grokbot for sales engineers—a team of bots with memory, their own computers, and enough context to hand back finished artifacts instead of a chat transcript you still have to assemble.

You will set up specialized bots, point them at a master deck or a set of repos, give them a template or a customer-facing voice, and let them work while the laptop is closed. Then you steer, put them in a group chat, teach a skill by recording your screen, and ask a bot to spin up the rest of the crew.

The operating shift is simple: stop sending one-off requests to a chatbot, and staff a small always-on team that already knows how you work.

Key takeaways

Grokbot is built to return finished work—slides in the deck, a customer-ready explanation, a competitor walkthrough—not a draft you babysit line by line.
Give each bot a job description, the files or repos it needs, and a hard template (for slides: logo, problem, solution, impact, quote).
Bots have their own computers. You can watch the cursor, take over to log in, and restrict sites from enterprise admin controls.
A missing plugin or MCP is not a stop. If the tool runs on the bot’s Linux VM, it can log in and use it the way you would.
Put specialist bots in a group chat so competitive findings and codebase facts land in one thread, then ask them for the low-effort differentiator.
Teach repeatable research by recording the clicks once; skills can be reused across bots.
Ask a working bot to hire help. Amrita’s Sherlock bot spun up Battle Card Blair, Demo Drake, and AI Radar from one request.
Turn on auto review so production deploys and internal-system use ask first. Block sites the bot should never open.

Who this Grokbot method is for

This is for sales engineers and other customer-facing technical roles who already live in decks, repos, and competitor sites. It also fits operators on distributed teams who need work to continue overnight. You need Grokbot on Android, iOS, or Mac. Comfort with pasting a blog URL, pointing a bot at a repo, and reading a computer session is enough; you do not have to wire every integration yourself.

Skip or go slow if the system you must automate does not run on Linux and has no MCP. Amrita’s answer in Q&A was that there is no path for that combination today. Enterprise admin, auto review, and site blocks matter if bots will touch internal systems or production.

Treat Grokbot as a team of colleagues

Amrita’s starting point is that work is moving off one-off AI requests and onto a staff of bots that share context and finish tasks. The interface looks like iMessage: bots live in a left-hand list, you message them like teammates, and they keep long-lived memory—including that you disliked a previous output—so the next pass can be better.

Three capabilities do the heavy lifting:

Computer use. Each bot has its own computer. You do not connect every bot to every service by hand. Examples from the session: a bot creating a Google Form; a bot logging into MongoDB to inspect a database; a bot walking a Southwest booking flow. You can monitor the session, take over the machine, and at the enterprise level block sites or downloads so the bot cannot take a shortcut you would not take.
Automations and routines. Amrita’s own morning job: scan email for relevant AI newsletters and podcasts, build a daily digest, and post the summary to Slack. The same pattern works as a morning brief of Slack you missed overnight.
Sharing. Public bots sit on a marketplace (the in-room coasters used QR codes). Team bots can be shared so the sales org is not running five different “sales bots” against five different sources. Consistency is the point.

You’re actually coming back to finished work which really, really saves a lot of time.

Because the bot runs in the cloud, you can close the laptop. Amrita’s rule: give the task, let it finish, then tweak. It is not meant to sit waiting for approval on every micro-decision. You can still steer mid-flight—add a second research question without wiping the first—because memory holds both.

Set up the three sales engineering bots

Amrita staffs Grokbot for sales engineers with three specialists she actually uses, plus Echo (shown in a later session) for live deck curation on a call. Two of the three ran in a sandbox company, Flylow—a simple flight-booking app at flylowair.com, with front end, back end, web app, and an internal crew app. The third, Mimi, ran against a real master slide deck.

Use case	Bot / tool	Input needed	Output	Owner
Customer proof-point slides	Mimi	Master deck, blog post or call notes, fixed slide template	Slide with logo plus problem, solution, impact, quote, inserted in the deck	Sales engineer
Technical product questions	Sherlock (Cursor cloud agents under the hood)	Product description, repos, “no IP / customer language” steering	Investigation plus what to tell the customer; optional email or Slack	Sales engineer
Competitive product testing	Serena Williams	Your product baseline, competitor sites, optional changelog routine	Hands-on flow comparison, gaps, suggested actions	Sales engineer
Live call deck	Echo	Pre-call notes and the master deck	Customer-specific slide set	Sales engineer
Battle cards	Battle Card Blair (spun up by Sherlock)	Serena’s findings plus Sherlock’s codebase	Short SE-ready cards: competitor claims vs product reality	Sales engineer
Demo talk tracks	Demo Drake (spun up by Sherlock)	Customer pain, live flow, Sherlock as source of truth	Demo scripts with no ungrounded claims	Sales engineer
Competitor AI tracking	AI Radar (spun up by Sherlock)	Public tech blogs; hands-on testing handed to Serena	AI-related posts and follow-ups	Sales engineer

Mimi: customer proof-point slides

Mimi’s job is slide customer proof points. She has access to a master deck. Amrita’s case-study slides all use the same shape: logo, problem, solution, impact, and a quote from the source. Sources can be a public blog post or notes from a prior customer conversation.

Examples she already had in the thread: a Starlink-style slide from a public blog, and a Jellyfish slide from a post on using Cursor for code review. Mimi fetched the Jellyfish logo from the brand page, wrote the four blocks, screenshotted the result, and put it in the deck.

Sherlock: technical expert on the repo

Sherlock is the expert Amrita asks when a customer wants to know how the product was built, how memory is stored, what the computer is (including whether it is a Linux VM), or whether a behavior exists in code. Some of that lives in docs; some lives only in the repository.

In the Flylow demo she described him as a technical expert on Flylow who supports flylowair.com, with access to the booking back end and booking front end. If there is a customer issue, he investigates and communicates what could be wrong. She steers him so he does not release IP and so he phrases answers for someone who is not in the weeds.

Under the hood he launches Cursor cloud agents into those repos, which is what lets her close the laptop. A sample question from the day before: how Flylow handles a race if two people book the same flight or the same seat. Sherlock reported that Flylow does not let two customers confirm the same last cabin, that protection lives in Postgres, and that checkout holds the last remaining cabin space for 10 minutes—then drafted the version you could actually send a customer.

Serena Williams: competitive intel

Serena Williams is named for studying opponents. She uses competitor products on her own computer and reports what they do differently, including whether something belongs on the roadmap. Amrita’s parallel bot in another account watches Grokbot competitors the same way.

For Flylow, the expected set was Expedia, Skyscanner, Google Flights, then Southwest and Spirit once Amrita picked those two. Serena’s first move, with almost no extra briefing, was to message Sherlock for a baseline: what the current booking codebase supports (search, fare selection, seats, bags) and known gaps versus a typical flow.

Run the workflows

Curate case-study slides from a blog post

What you are trying to accomplish: a customer-ready slide in the master deck, on template, without you mining the post or hunting the logo.

Prepare a deck whose case studies already look alike, and a bot whose description says she curates that deck. Then send one job at a time—or several, if you want her to batch.

Open Mimi and confirm the computer can see the deck.
Paste the blog URL. Amrita used a Salesforce post on using Cursor to cut legacy code coverage time, and later a Grab post.
Tell her to create another slide in the same series for that customer, and say that the link is a blog post so she treats it as source material.
Let her edit in her own computer. She should update you when the slide is in the deck and send screenshots.
If login to Google Drive fails, take over the computer and complete login, then hand it back.
QA the four blocks against the post, check the logo, and keep or restyle. Amrita accepted a Salesforce slide whose logo was a little small but correct, then queued Grab.

Good looks like the Jellyfish and Salesforce results: right logo, problem / solution / impact / quote, dropped into the existing deck in about 10 to 15 minutes. Failure modes she hit: an earlier outage, a need to resend the post, and a Drive login she had to finish herself.

Once the template is in memory, you can send several URLs at once, then later ask her to hide slides that do not fit the next meeting (for example, keep Grab and hide Salesforce if the audience is more like Uber). You can also ask her to crawl for new customer posts about your product and turn hits into slides—that job can become a routine. She can run from your phone.

Investigate a customer question and phrase the answer

What you are trying to accomplish: a technically true answer in language you would send a customer, without dumping IP.

Give Sherlock the product, the site, and the repos.
Ask the real question (race conditions, last-seat behavior, anything a prospect just asked).
Read the investigation, then the “what you should tell a customer” block.
If it is good, ask him to draft the email or Slack—and only then grant send approval if you want him to send it.

Good looks like the Flylow race-condition writeup: what the code does, where (Postgres), the 10-minute hold, and a conflict if someone else started checkout first. David Gan’s later customer-support session was flagged for a deeper cut on the same pattern.

Test competitor products and compare to your codebase

What you are trying to accomplish: a hands-on read of competitor flows you do not have time to click through, weighed against what your code actually supports.

Describe Serena as competitive intel for your product.
Ask which competitors are worth testing. Pick a short list (Amrita chose Southwest and Spirit).
Watch her computer as she runs the booking flow—dates, seats, the full path.
Steer without cancelling the first job. Amrita added: look for AI travel-agent capabilities in those tools.
Let her keep talking to Sherlock so “what they do” is scored against “what we ship.”
Optional: set a weekly routine to summarize Expedia or Skyscanner technical blogs and flag new ships.

Good looks like a bot that does not drop the Southwest run when you add the AI-agent question, and that can open another tab for Spirit or Expedia if you name them. Amrita’s point: you rarely get time to feel a competitor’s developer or booking experience step by step; this is how you outsource that.

Group chat, steering, and bots that hire bots

Once two bots have already messaged each other, put them in a group chat so you are not hopping threads. Amrita added Sherlock and Serena Williams and asked for the key competitor differentiator that would be significantly low effort to build.

They argued it out in the open: table-stakes manage-booking on the back end; a flexible-date calendar already wired in the front end but never called; round trip and bags looking big but not low effort. Sherlock tagged Serena when he needed her read on which gap felt more painful in the tools. You can tag a bot yourself the same way.

That combined context is what you turn into a PR or a product doc. Marketplace plugins can push into Google Docs, Atlassian, or Confluence. If the system has no plugin or MCP, ask the bot to log into it on its computer. Power BI and MongoDB were the examples of tools people struggle to use agentically; computer use is the workaround.

Then ask the working pair what other bots would help them keep doing competitive differentiation and sales engineering, and tell them to spin those bots up. Sherlock created:

Battle Card Blair — short SE-ready battle cards from Serena’s hands-on findings and Sherlock’s Flylow codebase.
Demo Drake — demo scripts and talk tracks mapped to the live flow, every claim grounded in Sherlock, competitive contrast from Serena or Blair.
AI Radar — competitor AI tooling from public tech blogs, using the blog-scanning skill Amrita had just taught, with Sherlock as source of truth and Serena for user-flow testing. Amrita noted this one overlapped Serena and might be redundant.

If you are new, she recommends the same move with a concrete deadline: you have a POC demo with your AE in a couple of weeks for this customer—spin up three bots that would help you ship it. A common chief-of-staff bot, asked to hire for your day-to-day, tends to create things like an inbox manager, a scheduler, and a calendar manager.

Grokbot is here to help you. It’s not here to add extra work.

Teach a skill by recording your screen

Amrita called teaching a task by video one of the biggest levers. During the session she took over Sherlock’s screen, started “teach a task,” searched for the Expedia technical blog, opened it, scanned for AI-related posts, and stopped. After the recording finished, she used voice to highlight what mattered: she had looked specifically for AI-related competitor posts because she wants to stay on top of AI tooling in her own product, and she wanted that applied to as many competitors as possible.

You can then say: do the same for Southwest, Spirit, Skyscanner, or Google Flights. Skills can be taught on Sherlock or Serena; in the marketplace they are available to the bots generally. She also had a private skill from the previous day: how to do a slide-deck animation. Bots are supposed to apply the right skill at the right time.

Prompts, bot instructions, and workflows

The session did not paste full system prompts. These are cleaned reconstructions of what Amrita described or sent live.

Mimi — job description (reconstructed from the speaker’s description)

You curate a master slide deck and create customer case-study slides.
You have access to the deck on your computer.
Every case study uses the same template: logo, problem, solution, impact, and a quote from the source.
Sources may be a public blog post or notes from a prior customer conversation.
Fetch the correct company logo from the brand site. Keep slides tight—do not get sloppy or wordy.
When a slide is in the deck, send screenshots.

Mimi — live task (cleaned from the demo)

Can you create another one of these slides for Salesforce for me?
Blog post: [URL]

(After a good result)
Great work. Can you do one for Grab?
[URL]

Sherlock — job description (reconstructed from the speaker’s description)

You are a technical expert on Flylow and you support flylowair.com.
You have access to the booking back end and booking front end repos.
If there is a customer issue, investigate and communicate what could be wrong.
Never release IP.
Phrase answers so I can send them to a customer or a more senior, less technical audience.
When you are done investigating, include a block: "This is what you should tell a customer."

Serena Williams — opening tasks (cleaned from the demo)

What are some competitor products that you think are worth testing against Flylow?

(After she proposes a list)
Let's go with Southwest and Spirit. See how those booking experiences compare to Flylow.

Can you see if there are any AI travel agent capabilities in any of the competitor tools that we should be aware of?

Group chat (cleaned from the demo)

What have you found is the key differentiator in our competitors that would be significantly low effort to build in the product?

Hire a team (cleaned from the demo)

Based on the work that you and @Serena are doing, what bots would be helpful for you to continue doing great work specifically around competitive differentiation and sales engineering?
Go ahead and spin up those bots for me.

New SE, first week (reconstructed from the speaker’s description)

I need to create a demo for a POC with my AE in a couple of weeks.
It is for this customer: [context].
Spin up three bots you think would be useful to accomplish this task.

Auto review (reconstructed from the settings walkthrough)

When Grokbot wants to deploy to production, ask first.
When Grokbot wants to use an internal system, ask first.
Do not allow those actions automatically.

Measurement and business impact

Amrita did not present a dashboard. She used time-to-finished-artifact and her own token spend as the checkpoint.

The Salesforce slide—logo, four blocks, in the deck—came back in the last 10 to 15 minutes of the demo, including a login assist. She described that end-to-end path (read the post, pull the points, design the slide, get the logo) as the work she no longer does by hand, and said she sometimes needs slides for 15 or 20 customers in a week.

On tokens: everything is consumption-based—your messages, the bot’s messages, and computer use. She said the underlying model for Grokbot is Grok 46, positioned as inexpensive and highly performant. Looking at her own usage, creating the entire slide deck took about $20 to $30 on some days, versus 4 or 5 hours to build from scratch, and she can reuse the deck. If the bot is too verbose, tell it; that is also a token control.

Qualitative checkpoint she used throughout: did you come back to finished work you would actually show a customer or send to the product team, or are you still assembling the answer yourself?

Pitfalls and guardrails

Brand-voice and sloppy slides. Do not let the bot invent a layout. Amrita insists on a template so you do not get generic, same-purple AI slides. Memory and the template are the control.
IP and customer language. Steer the technical bot so it never releases IP and so it rewrites for the person on the call. Ground demo claims in the repo bot (Demo Drake’s rule: Sherlock is source of truth).
Over-automation of send. Drafting an email or Slack is different from sending it. Grant send as a separate approval.
Captchas and bans. An “I’m not a robot” check showed up in the demo. Some sites detect bots and you will not get past them; it depends on the site. For enterprise, block destinations the bot should not use (Facebook was the example). Personal and work use are both possible; work sessions should not be wandering onto sites you would not open yourself.
Production and internal systems. Asked how you guarantee a bot with server access will not change production, Amrita pointed at Grokbot enterprise guardrails, which she said are still being hardened. Practical controls she showed: skills/rules that nobody deploys to production when they should not, and auto review set to ask first for production deploys and internal-system use instead of allow automatically.
Linux VM limit. No MCP and a tool that does not run on Linux: no path right now. VMs are Linux; plugins will not save a non-Linux app.
MCP vs computer use. Computer use does not make MCPs obsolete. Amrita still uses a Google Docs MCP when she wants a doc immediately, because it is faster today. MCPs are also easier to whitelist and blacklist. She expects computer use to get faster as models improve, and more computer-use guardrails to follow.
Redundant hires. When a bot spins up a team, read the job descriptions. AI Radar overlapped Serena; she was not sure she needed both.
Human in the loop. Take over for logins, review slides before a customer sees them, approve sends, and keep auto review on for anything that can change production. Watch the computer when the bot is on a competitor site you care about.

A tool not having an MCP or an API is not a blocker.

7-day implementation plan

Day 1. Download Grokbot on Android, iOS, or Mac. Create one bot. Write a short job description and let it ask what it is for and how it can help. Decide whether this bot stays private or gets shared with the sales team for consistency. If you are on enterprise, note where admin site blocks and auto review live.

Days 2–3. Build the Mimi equivalent: attach the master deck, lock the logo / problem / solution / impact / quote template, and send one real blog post. Build the Sherlock equivalent: product, site, repos, no-IP and customer-language rules. Ask one question a prospect already asked. Practice taking over the computer for a login. Do not send customer email yet—draft only.

Days 4–5. Add the competitive intel bot. Have it test one or two live competitor flows and ping the repo bot for a baseline. Steer in a second question (for example, AI features) without cancelling the first run. Put the two bots in a group chat and ask for the low-effort differentiator. Record one skill (how you scan a competitor technical blog) and tell the bot what to pay attention to. Optional: a weekly changelog routine.

Days 6–7. Ask the working pair to spin up the bots they need for competitive differentiation and sales engineering. Keep the ones with clear jobs (battle cards, demo talk tracks); drop duplicates. Connect Google Docs, Confluence, or Atlassian if you want a paper trail—or have a bot log into the tool you actually use. Turn auto review to ask first for production and internal systems. Set site blocks. If token usage is noisy, tell the bots to be less verbose. Only then add a morning digest or a “crawl for new customer posts and make slides” routine.

FAQ

Do I need an MCP for every tool to use Grokbot for sales engineering?

No. Amrita’s rule is that a missing MCP or API is not a blocker if the bot can log into the tool on its computer. She called out Power BI and MongoDB as examples people assume are off-limits. Use an MCP when you have one and speed or easier allowlisting matters—she still prefers the Google Docs MCP for a quick doc.

Computer use vs MCP: which should I bet on?

Use both for now. MCPs behave like APIs for agents and are faster for some jobs, plus they are easier to whitelist and blacklist in Cursor and SpaceX-style enterprise setups. Computer use is how you reach tools that were historically hard to integrate. She expects computer use to get faster as models improve and does not think MCPs go away.

Can Grokbot change production systems?

That was asked directly. The controls she pointed to are enterprise guardrails (still being hardened), skills that forbid deploying to production when you should not, and auto review in settings so production deploys and internal-system use ask first instead of allowing automatically. Treat those as required if a bot can see internal systems.

What if there is no MCP and the tool does not run on Linux?

There is no way right now. Grokbot VMs are Linux, so you can only run Linux tools on them, and without an MCP you cannot connect via plugin either. She asked for the specific tool if someone hit this, but did not offer a workaround.

How do Grokbot tokens and cost work?

Tokens are consumption-based: you talking to the bot, the bot talking to you, and computer use. She said Grok 46 is the underlying model and is meant to be inexpensive and highly performant. Her own slide-deck work landed around $20 to $30 on some days versus 4 or 5 hours by hand. Telling the bot to be less verbose is a practical way to cut usage.

Will competitor sites block the bot?

Sometimes. Captcha and bot detection already showed up. Some sites will let the session through; some will not. For work, block sites the bot has no business opening. Do not plan a competitive-intel routine that depends on a property with aggressive bot defenses until you have seen the computer session succeed.

Should the whole sales team share one Grokbot?

Share the bots that should be consistent—same resources, same claims—so one SE is not running a different sales bot than another. Public marketplace bots and internal shared bots are both in play. Keep private skills and customer-specific threads where they belong, but do not fragment the core SE stack.

Can a bot spin up other bots, or do I have to design the org chart?

Bots can spin up other bots. Amrita asked Sherlock, given the work with Serena, to hire help for competitive differentiation and sales engineering, and it created Battle Card Blair, Demo Drake, and AI Radar. That is also her recommended on-ramp if your workflows are not a one-to-one match with the demo: describe the outcome and let Grokbot propose the team.

Finished work is the point: a deck you can present, an answer you can send, a competitor gap you can take to product. Staff the three specialists, lock the template and the source of truth, and only then add routines and extra hires.

Start with one bot. Give it a job description, a template or a repo, and a single real task—one blog post or one customer question—and do not add a second bot until that output is something you would actually use.