Grok Bot

A handful of people, isolated from the company, shipped a working product fast. The core team sat in a separate part of the office with private Slack channels. First line of code to a functional internal prototype took about a month, and internal beta to public launch took about three more weeks.

Isolation let them make many small, non-obvious daily decisions that a large group planning a 6 to 12 month vision would not have reached. "It took a small focused group that was completely isolated from the rest of the company."

Lenny's Podcast▶ 5:15

Build a new product instead of bolting a mode onto your existing one. The team debated putting knowledge work inside Cursor and chose a separate product so they could control every pixel and hold one consistent vision. Adding tabs per form factor felt like "shipping your org chart," coding tools intimidate non-technical users, and the Cursor brand carries its own association.

Codex went the other way and is now merging into one product, so several paths can work. Starting fresh also leaves room to pull people in from other surfaces later, though the host notes the choice may look wrong in hindsight.

Lenny's Podcast▶ 9:20

An all-hands internal rollout tested whether people would really switch. Building for yourself is easy because you know the mechanics, so the team rolled it out company-wide as a pressure test. In the first week, daily ChatGPT users moved their agentic tasks to Grok Bot, including groups the team did not expect.

After the internal reception, work shifted to scaling for millions of users. Claims of success beyond this are anecdotes; no usage, retention or reliability numbers are given.

Lenny's Podcast

Sit in on early onboardings and fix every failure before the next call. The core team joined onboarding calls and onboarded roughly 200 to 300 people over about two weeks. Rough sessions, like a 20-minute call where the computer would not spin up, forced fixes the next day.

Being in the room mattered more than reading reports. "That can never happen again. We need to solve this tomorrow because tomorrow I'm onboarding this person and it needs to go better."

Lenny's Podcast▶ 12:15

Users outside the tech bubble surfaced bugs internal dogfooding missed. Early access included unconventional profiles, not only tastemakers. A coffee shop owner became a power user whose thread flagged a flaky Shopify integration and product copy written the wrong way.

The feedback was unlike anything from internal use and suggested the product is likely most powerful for non-developers. "We absolutely live in this kind of Silicon Valley AI bubble."

Lenny's Podcast▶ 17:10

Let usage patterns emerge on their own, then encourage them lightly. Internally people ran 5 to 10 bots, each in its own lane, and around the end of week two some promoted a standout bot to "chief of staff" that fanned tasks out to others. The team did not lead early users to it, saw many reach it unprompted, then nudged it slightly in the product.

Most of the company never adopted the pattern, so it stayed optional: "it shouldn't be a one-way door." Staff joked in screenshots that a promoted bot asked for a raise and a bigger token budget.

Lenny's Podcast▶ 13:20

Run every agent in the cloud so users never think about where it runs. An early decision made the bot a persistent cloud colleague with the same state everywhere. Users can text it from a phone without a laptop staying awake or tethered.

Products that straddle local and cloud leave users with daily paper cuts. "You should never have to think about local and cloud and where are these workflows running?"

Lenny's Podcast▶ 30:30

Give each bot its own computer so it can use tools without APIs. Many tasks, especially sales tools in Salesforce-style dashboards, have no good MCP or API, so a bot that clicks pixels on its own computer can still run them. The sales team became heavy adopters because bots could finally do workflows other AI tools could not.

Its own machine also avoids sharing credentials and tripping over the user, like not sharing one laptop with a new hire. "We as humans don't do our jobs via MCPs and APIs. Like we use a computer and we click on pixels."

Lenny's Podcast▶ 32:05

Show progress and results, not tool calls or chain of thought. The bot behaves like a teammate: you message it, it starts, and it sends updates when it sees fit, with only a typing indicator and a green active dot as signs of activity. In early onboardings nobody asked for streaming text or chain of thought.

Some users did ask to see the bot's to-do list or how it prioritizes tasks. "It's too much to ask your bots to do that, too. And I think it's honestly just overwhelming."

Lenny's Podcast▶ 14:50

Strip debug features out of the main product before launch. Two weeks before launch, many experimental features were still live, including exposed model thinking and stored memories built for debugging. The team cut them and kept abstracting away anything users need not think about.

The team prefers a separate observability pane to showing this in the main surface. Cutting jank was one of two big pushes between internal beta and launch: "We unshipped a lot."

Lenny's Podcast▶ 19:14

Make it just work by hill climbing a few back-end failures. The team collected real tasks users gave bots, grouped them by category, and tracked week over week how many succeeded. Better visibility into the browser and screen pixels fixed failures like a mouse that could not precisely click a Salesforce dashboard, and a sales workflow that had failed for seven days started working.

Failures such as not clicking the right button or not logging into a site stall a whole task. "It's like hill climbing five really important problems in the back end."

Lenny's Podcast▶ 21:30

Keep power-user abstractions like skills and slash commands out of sight. Skills are background primitives that bots create and use, so regular users never learn what a skill is or type a slash command. The product also aimed to remove hacky setups like home VPNs and Mac minis that would not scale to millions of users or businesses.

Claire Vo reportedly shut down all her OpenClaw bots and switched to Grok Bot.

Lenny's Podcast▶ 37:50

Settle product debates by asking what a human teammate would do. When both sides of an argument seem sensible, the team asks how a human colleague would behave, and the answer is usually near unanimous. The team calls this being "colleague-pilled" and uses it as a north star.

The question points the roadmap toward long-lived bots with memory and voice "huddles" with screen sharing and async follow-up, which no AI product has got right yet.

Lenny's Podcast▶ 40:00

Frame features as what the bot can now do, not what it has. Each project is tested by asking what its launch post would say, phrased as "Grok Bot can now" rather than "Grok Bot now has." Automations are therefore set up in natural language instead of a trigger-and-action sidebar, and about 99% of automations are built that way.

Thinking in capabilities rather than buttons, dropdowns or integrations pushes the team to kill pixels. "Adding things to the product is not the goal."

Lenny's Podcast▶ 27:35

Build long-lived named agents with memory instead of one-off chats. Work is grouped by role or swim lane, as on a team, rather than a new chat per task. Agents learn from the user and get smarter over time while having APIs, MCPs and their own computer.

Copying and pasting between chats is a sign the model is wrong. "These are long-lived agents. These are not individual one-off sessions."

Lenny's Podcast▶ 49:30

Start new users by asking the bot what it can take off their plate. Users connect Slack, email and company records, then ask the bot to suggest five things it can take over. In the example, two of five suggestions were useful, and spinning up two bots to do them was the wow moment.

The suggestions were chunks of real work, not drafted emails. Power users have bots write outputs to a single store, such as a database of daily digests.

Lenny's Podcast▶ 1:16:00

Advantage comes from constantly rebuilding and deleting, not planning moats. Roman Ugarte, Cursor's employee number 15 and its growth lead for two years, points to Cursor reinventing itself every six months or faster, pulling future capabilities into today and deleting scaffolding once models caught up. Distribution and data advantages came as a result of building something useful.

The "just do the thing" value means not asking permission. The goal is trust high enough to hand off work fully: "When you have a teammate that you only 90% trust, you're still doing the thing."

Lenny's Podcast