Tag
A study finds that 58% of tokens consumed in failed AI agent runs occur after the model had a clear signal to stop, highlighting a missed opportunity for mid-run cost savings.
The article analyzes a 2026 paper by Bai et al. showing that subagents and context bloat cause token costs in long agent runs to be ~1000x higher than chat, and presents three practical fixes (PLAN.md, read budget, out-of-band notes) that reduce token usage by 70-90%.
An AI researcher ran 13 controlled experiments on a multi-agent coding system, finding that dependency-ordered coordination significantly improved success rates while persona backstories had no measurable benefit.
A developer tool that records AI agent runs inside a sandboxed GitHub repository, capturing terminal/browser sessions and turning them into replayable narrated videos for improved observability.