A DoorDash engineer shares his personal journey from using Cursor for AI-assisted coding to mastering long-running autonomous agents with Claude Code. After failed experiments with multi-agent teams (which produced inconsistent, unusable output due to lack of shared context), he discovered the Research-Plan-Implement (RPI) framework, the agent loop pattern, and antagonistic reviewers in fresh context windows. These insights led him to build Agentic Orchestrator — an open-source TUI state machine that drives features through phases (knowledge base → inquiry → research → design → roadmap → planning → implementation → review) with git worktree isolation, crash recovery, and deterministic phase transitions. The post also candidly addresses the cognitive toll of high-density agentic workflows, warning that AI amplifies work density rather than reducing hours, and offers practical advice for developers starting with autonomous coding agents.
