Meet Grok Bot: Your Team of AI Agents

This session introduces GrokBot, described as a team of always-on agents available at your disposal. Key features highlighted include persistent memory, access to tools, and the ability to communicate with each other. The core shift in thinking with GrokBot is moving from a task-based model to a persona-based model, where bots act as colleagues with specific roles and expertise. The presentation demonstrates GrokBot's capabilities through two use cases: a general-purpose organizational assistant and a more engineering-focused application. **General Purpose Use Case: Chief of Staff Bot** The first demo focuses on a "Chief of Staff" bot, designed to help organize a user's day. This bot is configured to work with three other specialized agents: an Inbox Manager, a Calendar Scheduler, and a To-Do List Organizer. * **Bot Creation and Configuration:** Users can create bots, assign them roles (e.g., "day-to-day organizer"), and then instruct them to spin up other agents. Crucially, bots can be granted access to user tools and plugins like Gmail and Google Calendar. The interface is designed to be familiar, resembling a chat application. * **Inter-Bot Communication:** The Chief of Staff bot has a group chat with the other agents, enabling them to communicate and collaborate. Users can also directly interact with the Chief of Staff bot, delegating tasks. * **Task Delegation and Scheduling:** The presenter demonstrates delegating tasks by speaking them into voice mode (e.g., "do laundry," "call grandmother," "buy groceries," "help friend with presentation"). The Chief of Staff bot then coordinates with the Calendar Scheduler and To-Do List Organizer to slot these tasks into the user's schedule. * **Learning and Skill Creation:** GrokBot agents learn from user preferences. For example, the presenter teaches the Chief of Staff that laundry should be scheduled late at night by creating a "skill." This skill, representing a superpower for the agent, is then permanently integrated. This learning process extends to agents having their own "computers" (remote VMs) where they can be taught new tasks by observing user actions, such as booking a flight. This demonstration of "teaching a task" transforms a recorded user action into a reusable skill. * **Computer Use:** GrokBot agents have access to their own computers, allowing them to perform tasks that might not be directly covered by existing plugins. This includes actions like creating presentations in PowerPoint or Google Slides, by teaching the agent how to use the application. * **Remote Operation:** A significant benefit highlighted is that GrokBot operates completely remotely. Agents continue to run even when the user's laptop is closed or they are away, saving time and allowing for background task completion. * **Routines:** GrokBot supports creating "routines," which are scheduled or recurring sets of tasks. An example is an "inbox briefing routine" where the Inbox Manager is instructed to go through newsletters daily, extract highlights, and send them via Slack. This routine can be triggered by time or other events. * **Memory and Customization:** The persistent memory stored in S3 buckets allows agents to recall past interactions over extended periods. Users can improve agent performance by creating skills, routines, and defining anti-patterns (what the agent should *not* do). * **Controls and Guardrails:** Users can set system-level controls to manage bot actions. This includes requiring explicit permission before certain actions are performed (e.g., sending emails, adding calendar events to shared calendars). GrokBot also prompts users during initial setup to establish these guardrails. * **Engineering Use Case: Tech Lead and On-Call Agents** The second demo shifts to an engineering context, showcasing how GrokBot can be integrated into software development workflows. * **Engineering Personas:** The setup includes persona-based agents like "Tech Lead Taylor," "Backend Bobby," "Frontend Faye," and "QA Quincy." These agents have access to the project's repository (e.g., a flight booking system called Flylo). The key is that these personas are adaptable; Backend Bobby can work on different repositories because of his role-based expertise. * **Feature Development:** The presenter asks Tech Lead Taylor to work with the engineering team to build a new feature (adding lodging and hotel booking to a flight website). This involves spinning up new endpoints, UI elements, and routes. QA Quincy is included to perform end-to-end testing, demonstrating the ability to record screen activity, including network inspector details. * **Agent Orchestration and Collaboration:** The system highlights seamless orchestration and coordination between engineering agents. For instance, Backend Bobby and Frontend Faye can communicate directly about API contracts and progress, minimizing waiting time. Agents can also adapt to team changes, like looping in others if a team member is unavailable. The interaction shows Faye waiting for Bobby's API contract and then immediately starting frontend work once it's provided. * **On-Call Agent:** An "On-Call Ollie" agent is created to monitor Datadog and PagerDuty for API error spikes or customer incidents. This agent can be triggered automatically to handle issues, preventing disruption to human engineers. The agent proactively suggests creating automations for PagerDuty events and polling Datadog, seeking user permission before proceeding. * **Computer Use in Engineering:** QA Quincy's ability to record full-stack tests, including network inspector data, exemplifies the importance of agents having access to their own computers. This allows for comprehensive testing and validation. * **Code Review Integration:** While agents handle task orchestration and initial implementation, the actual code review process can still be done via tools like Cursor or GitHub. The agents can generate pull requests (PRs) that users can then review and merge. The demo shows how QA Quincy adds video recordings of tests as comments to PRs, providing detailed evidence of functionality. * **Product Page Agent for Release Notes:** A "Product Page" bot is introduced to take details from completed engineering tasks and generate customer-facing release notes. This demonstrates how agents can collaborate across departments (engineering, product, marketing) to communicate product updates. * **Mobile App:** GrokBot is also available on a mobile app, offering a similar interface and functionality to the desktop version. This allows users to interact with their agents, view conversations, and use voice mode on the go. * **Broader Use Cases:** The presentation briefly touches upon other potential use cases, including marketing (media manager, growth manager), sales (prospecting bot, outbound bot), and finance/RevOps. Upcoming workshops will delve deeper into these specific applications. * **Key Differentiators Recap:** 1. **Teammate-like Bots:** Bots act as colleagues, communicating with users and each other. 2. **Persistent Memory and Continuous Learning:** Bots remember past interactions and improve over time. 3. **Continuous Operation:** Work continues even when the laptop is closed. * **Feedback and Access:** Users are encouraged to sign up at x.ai/bot for Cursor Ultra or Super Grok Heavy accounts to access GrokBot. Feedback is highly valued. The session concludes with a Q&A, addressing questions about self-improvement loops, bot memory (comparing it to human memory and LLM capabilities), the "cold start" problem for new tasks, passive workflow discovery, and the potential for advanced voice interactions. The presenter emphasizes that while current memory is more LLM-like, the goal is to evolve towards more human-like episodic and conceptual memory. The importance of providing context and teaching skills remains crucial for optimal performance. The possibility of integrating with iMessage for texting and the absence of a context window limit due to S3 memory storage are also noted. The session highlights the fun and innovative nature of GrokBot, positioning it as a significant leap in AI technology.