The Rise Of AI Coworkers

Summary

Grok Bot is presented as a platform for creating AI teammates—bots with names, roles, skills, tools, and their own identities. Its design is built around bots rather than chats or files. The team’s goal is for bots to use the same workplace tools as people and stay active in the background, proactively taking on work or surfacing useful information.

How Companies Are Using AI Teammates

• Start small, then expand trust: A common approach is to begin with a limited, read-only task, then grant more access as the bot demonstrates reliability. The longer-term aim is for bots to own larger responsibilities or outcomes.
• Create bots around work: Users can organize bots as specialists for particular functions or customers, or use fewer, broader bots. The guests say dedicated bots focused on specific problems can be easier to manage than one bot expected to handle everything.
• Share expertise across a team: Grok Bot’s Team Bots feature lets a domain expert build a bot and share it with colleagues. Examples include marketing bots working with design bots to create and test ad assets, support bots monitoring tickets, sales bots researching leads, and data bots answering questions using read access to company data.
• Put bots where work happens: Slack and X can bring bots into existing discussions, giving them context and enabling collaboration with human coworkers. The team aims to work inside existing business tools rather than rebuild them.

Identity, Access, and Collaboration

The guests describe bot identity and permissions as unresolved industry-wide challenges. Their proposed starting point is to apply familiar employment principles: give a bot credentials appropriate to its role and experience, then increase its access as it earns trust.

Bots can also collaborate with one another. Grok Bot provides a way to view their conversations, and the guests describe examples of bots coordinating through Slack. They note that multi-agent work still needs better handling of authority and social context—for example, knowing whose instructions to prioritize in a team discussion.

Coding Workflow and Quality Controls

Lauren describes using a bot in a Slack thread to turn a feature idea into a working implementation. The bot launched a Cursor cloud agent with its own virtual machine, produced screenshots and videos for feedback, and completed the feature through back-and-forth collaboration in the thread. Lauren says the team shipped that feature without her opening Cursor or reviewing the pull request directly.

She attributes this kind of workflow to investment in skills, codebase guardrails, and automated checks. Her engineering output had grown to more than 4,000 pull requests a month, but the emphasis is not simply on producing more code. The codebase should encode standards so agents can catch common problems themselves.

A key practice is to close the feedback loop: teach the agent to run the software, reproduce issues, and check its output against the desired behavior. Once that works reliably, alerts or other events can trigger agents to investigate issues and act without constant supervision. Lauren also describes using project-coordinator agents to manage subagents.

Lauren’s PAC (“Potato Stack”) is an open-source collection of engineering skills intended to teach agents to work rigorously in a style she trusts. She encourages engineers to adapt tools and skills to their own workflows.

Reviewing Code and Reducing Token Costs

Lauren recommends putting most quality controls into the codebase and automated tooling rather than relying on increasingly expensive agent reviews. Her “inverted pyramid” approach prioritizes:

1. Codebase architecture and conventions
2. Static analysis, lint rules, compilers, and other automated checks
3. Agent instructions and reusable skills
4. Human review when appropriate, especially for sensitive infrastructure or deployment work

She cautions that adversarial or multi-model reviews can produce excessive or irrelevant feedback when agents lack the product context or judgment to distinguish consequential issues from theoretical ones. As an example of a durable guardrail, Grok Bot’s Electron app enforces rules against importing between main-process and renderer code, helping avoid performance problems without repeatedly spending review effort on the same concern.

What Changes as Execution Gets Cheaper

The speakers argue that AI makes fast, inexpensive, high-quality execution more attainable, but shifts the bottleneck toward choosing the right goals and defining them clearly. When teams can build many ideas quickly, product direction, alignment, and judgment become more important.

They distinguish routine maintenance—such as fixing small bugs based on customer feedback, which can often be automated—from changes to core user experiences, which need careful internal testing and a high bar before shipping. Human taste and attention to user experience remain important.

The conversation also considers the idea of AI-native companies, where people may spend less time executing tasks and more time defining outcomes, coordinating work, and helping colleagues use AI effectively. Lauren’s account of moving from finance into engineering highlights the value of building projects, sharing what you learn, and communicating your ideas.

Key Workflow Guidance

• Give bots narrow, low-risk access first; expand it as they earn trust.
• Share specialist bots across teams to spread domain knowledge.
• Use existing workspaces such as Slack to provide context and enable collaboration.
• Build automated tests and codebase constraints so agents can check their own work.
• Reserve human attention for product judgment and high-risk decisions.

Main Speakers and Sources

• Lauren Tan — Grok Bot team; discusses AI-assisted engineering workflows, PAC, and code-quality practices.
• Roman Ugarte — Grok Bot team; discusses the product, team bots, and company use cases.
• YC Lightcone hosts — Interviewers and commentators; one host is addressed as Gary.