GrokBot developers scale output with autonomous AI teammates

Lauren Tan and Roman Ugarte developed GrokBot, a product that allows users to create AI teammates with their own identities, telephone numbers, and credit cards. The team incubated the product in seven weeks, moving from the first line of code to launch. The platform is integrated into the default direct messages on X and enables bots to operate with the same tools as humans, including access to Slack, email, and specific business applications. Internal use cases include a marketing team granting a bot access to Google and Facebook ads with a budget to A/B test assets, and a support team using bots to reproduce issues and triage tickets. The product recently launched 'team bots,' allowing shared bots with defined owners, managers, and collaborators. Small businesses are early adopters, using GrokBot for accounting and bookkeeping due to hiring difficulties. The system supports 'multiplayer' interactions where bots chat with each other in Slack threads, with humans able to observe these conversations. Tan argues that organizing work around bots rather than chats or files is the key operative principle for the next era of software, enabling non-engineers to automate complex functions.

Tan reports pushing over 4,000 pull requests per month, up from 2,000 previously, using a workflow centered on GrokBot and Cursor. Her process involves interacting with a bot named 'Potato' in Slack threads, where the bot spawns Cursor Cloud Agents with their own virtual machines to prototype features and return videos and screenshots. Tan describes an 'outer loop' where context is gathered directly from Slack rather than manually copied. She developed 'P-Stack,' an open-source set of skills inspired by G-Stack, which teaches agents to perform rigorous engineering work. P-Stack is installed in the company repository, with an internal version called 'Lauren mode.' Tan advocates for an 'inverted pyramid' approach to code quality, prioritizing investment in the codebase, static analysis, and lint rules over extensive token-consuming code reviews. For example, the GrokBot desktop app enforces a dependency graph checker to prevent long tasks on the Electron render thread, ensuring 60 FPS performance without manual review. Tan notes that agents can now merge their own pull requests, with humans focusing on product vision and intent specification rather than execution. She argues that the new role of a software engineer is to define intent and build guardrails into the codebase, allowing agents to handle execution and self-correction efficiently.