Descriptions:
Cole Medin extends Anthropic’s open-source long-running task harness into a full AI engineering system, connecting Claude Code to Linear for task management, GitHub for version control, and Slack for progress notifications — all orchestrated through the Claude Agents SDK with the Arcade platform providing MCP-based service integrations. The result is a coding agent that behaves more like a complete engineering team than a code completion tool.
The architecture separates concerns using dedicated subagents: an initializer agent reads the project spec (called an appspec, analogous to a PRD), scaffolds the codebase, creates the Linear project, and sets up the GitHub repository. Implementation agents then tackle individual Linear tasks in isolated context windows, committing each completed feature as a separate pull request and updating a meta progress-tracker task to hand off state to the next agent session. Subagent isolation is the explicit mechanism for preventing context bloat across a long multi-session build.
In the live demonstration, Medin builds a research dashboard — an agent-driven UI that dynamically generates data visualization components based on pasted research text — from scratch. The harness autonomously creates 44 Linear tasks, manages the full git workflow, and posts real-time Slack updates throughout. The video includes architecture diagrams, a quick-start guide for WSL, Mac, and Linux environments, and a public GitHub repository with the full harness code. Medin positions this subagent pattern as the practical path for pushing AI coding beyond what single-session agents can reliably accomplish.
📺 Source: Cole Medin · Published February 02, 2026
🏷️ Format: Hands On Build
