Headlong launches open-source persistent agent harness

Headlong makes agent persistence feel tangible, but its always-on design turns autonomy into an infrastructure and safety problem.

• –Continuous reasoning costs roughly $1–$2 per hour, even without user activity
• –Tiered context and DAG-based trajectories offer a practical approach to long-lived agent memory
• –Shared Slack, Telegram, and chat interactions create useful team context but eliminate strong privacy boundaries
• –Real shell execution and self-modification make sandboxing, spend caps, and service isolation essential
• –The project is compelling research software, though its benefits remain qualitatively demonstrated rather than benchmarked