Builder ships Agentic IDE, finds Cache Read dominates token use

The builder says existing Agentic IDEs such as orca and kiro are already good, but wanted something different: use of contracted subscriptions rather than APIs alone, coordination between those subscribed agents, quantitative measurement, freedom for AI inside a sandbox that does not touch the local environment, user-defined workflows, and AI-to-AI debate. Early development leaned on Claude Code as a single agent, and now the IDE develops itself.

The workflow view shows one row as a single workflow run and each box as a step run by an agent, with agents deciding pass or fail and sending failures back to earlier steps. Time and token measurement was a deliberate focus: which agent spends time and tokens where, how efficiently information passes between agents, and how much PC resource is used.

On the AI-to-AI debate experiment, the developer had models debate adopting Jev, Type Safe AI, with gemini 3.1 pro and ops5.5; depending on model character, they deny what is factually wrong, which raises confidence. Jev and Clef are planned as decision-specialized models, and local LLM integration is also wanted, though hardware procurement is proving difficult with NVIDIA expensive and AMD AI Pro 9700 or Intel B70 under consideration.