Reladraw: a diagram language built for you and your agents

Agents draw diagrams from relative placement, not coordinates — open source, npm, a Claude skill. Plus a study on how watermarking shifts agent tool-calls.

Nowline Sep 27 9:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Reladraw: relative placement, not coordinates

    A new open-source diagram language places elements with statements like “right of” and “below,” so an LLM agent can lay out an architecture or flow chart without burning tokens computing pixel coordinates. It runs in a browser playground, installs over npm, and ships a Claude skill you can hand an agent today — a clean weekend upgrade for anything that generates diagrams.

  • Watermarking quietly moves your agent's hands

    A Lasso Security study circulating this week finds text watermarking degrades tool-call accuracy on 6 of 7 open models and flips refusals: paired “churn” averaged 6.5% and jumped to 23.5% on gemma-3-27b once a prompt injection was added. If you deploy a watermarked model, re-run your agent evals with that exact key — behavior that looks stable on clean inputs may not hold under attack.