AI Agents Just Became Multiplayer. Here's How to Build Them.

Summary

Nate Herk begins by demonstrating an interface combining elements of Slack and ClickUp, showing a channel named ais-media where human team members collaborate directly with an AI agent named Herk. He explains that current AI workflows are predominantly single-player, meaning individuals assemble personal systems by connecting tools like Claude Code, Codex, and Grok Bot to private accounts in Slack, ClickUp, Notion, and Gmail. While this personal automation dramatically amplifies an individual's output, it leaves organizations fragmented, as teammates duplicate effort, maintain separate contexts, and manually reconcile data across shared spreadsheets and project trackers. To bridge this gap, Herk presents Kylon, a platform built for multiplayer AI team coordination.

Inside Kylon, users organize work into specific rooms and channels populated by both human colleagues and autonomous agents. Herk demonstrates a weekly commitments dashboard built directly into the media channel, which synchronizes with ClickUp to track project completion rates and automatically runs a daily check-in routine at five in the afternoon. In another channel dedicated to Instagram operations, Herk pairs his local Claude Code and Codex agents with scheduled tasks that scan sources like X and Bletato every six hours. The agents automatically research industry news, generate draft carousel posts, and populate an interactive Kanban board app inside the channel, allowing team members to review copy, inspect analytics, and approve publication.

Herk then walks through the configuration of new workspaces, starting with team naming, web address verification, and software integration. Kylon connects to external services including Gmail, Google Calendar, GitHub, Linear, Slack, and Outlook through OAuth or custom API keys, while also supporting Model Context Protocol servers. Team members can launch agents hosted directly on Kylon Cloud across three intelligence tiers, designated as Fast for summaries and basic editing at one-tenth max cost, Pro for standard daily operations at half max cost, and Max for complex engineering tasks using Claude 3.5 Opus. Each agent profile can be customized with specific personas, thought processes, and voice guidelines.

Alternatively, users can link external CLI subscriptions such as Claude Code, Codex, or Devin by pasting terminal connection prompts into their local developer environments. Running agents locally preserves existing tool variables, repositories, and credentials while bypassing Kylon workspace credit charges. Herk notes an operational tradeoff: locally hosted agents become unresponsive when the host machine is powered down, meaning always-on background availability requires continuous computer uptime, dedicated mini PCs, or virtual private servers. He concludes by examining Kylon's billing structure, which bills per workspace rather than per member seat, providing granular breakdowns of credit consumption across individual users, communication channels, and active agents.