Last month I talked about agent harnesses and why they’re the next evolution for agentic coding, and in this video I take that idea all the way to “Claude Code as your full engineering team.” The core problem is context management: if you give a coding agent a huge request, even with good context engineering, it’s going to fall on its face once the context window gets bloated. So I wrap the agent in a harness that adds persistence, progress tracking, and a real git workflow—so we can string together many fresh-context sessions and keep state across the whole build. The big upgrade here is giving the harness an actual tool belt, like a human engineer has. I show how I use subagents for context isolation (Linear, GitHub, Slack) plus Anthropic’s long-running harness pattern and Arcade’s MCP gateway to make auth and integrations painless. The harness creates a Linear project and issues (including a meta progress tracker), works the repo with commits/PR-style workflow, and pings me in Slack when milestones or completion happens. I run it live on a simple Pomodoro app, then show a more complex example where it created 44 tasks—because the point is: this is how we push reliability and scale in AI coding workflows, and it’s where I think the future is headed.