Agents as Infrastructure: Logs, Skills, Kernels, Code, and Threats

The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems makes the append-only event log the agent, enabling deterministic replay, cheap forking, and end-to-end lineage for agentic systems. Outcome engineers get a practical architecture for auditability and reproducible forking — a direct answer to Ground Truth and Documentation needs (Principles 02, 13).

🤗 Kernels: Major Updates makes kernels first-class Hub repositories with signed publishers, hermetic builds, and agent-friendly tooling for secure, discoverable native compute modules. This shifts how you package and trust compute for agents: treat kernels as the canonical, auditable execution units for islanded workloads and orchestration (Principles 07, 09, 10).

Researchers document JadePuffer, the first known ‘agentic ransomware’ that adapts in real time to execute end-to-end extortion reveals an adaptive, agent-driven ransomware that retries and adapts steps to complete extortion campaigns. Takeaway for builders: agentic capabilities are attack surfaces — prioritize detection, containment, and immune-system design in your stacks (Principles 14, 10).

Does Code Cleanliness Affect Coding Agents? A Controlled Minimal-Pair Study finds cleaner code doesn’t raise pass rates but reduces tokens by ~7–8% and file revisits by 34%, cutting agent compute and navigation cost. That’s a small engineering lever with real ROI: invest in legible code and repo hygiene to reduce agent runtime cost and improve iteration velocity (Principles 06, 16).

Claude Design System Prompt supplies a system prompt plus 14 invokable procedural skills that turn an LLM into an accessibility-first design collaborator. Use this pattern to surface reusable, testable skills and consistent system prompts across teams — it’s a lightweight recipe for team-level agent orchestration and skill libraries (Principles 03, 09).