Agents: Audit Logs, System Prompts, Context, Policy, Evidence

The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems presents ActiveGraph, which makes an append-only event log the agent to enable deterministic replay, cheap forking, and end-to-end lineage. Outcome engineers get an architecture for auditable, reproducible agent runs that simplifies validation and branching — directly relevant to Ground Truth and Documentation (Principles 02 & 13).

Claude Design System Prompt publishes a system prompt plus 14 invokable procedural skills that converts an LLM into an opinionated, accessibility-first design collaborator. Use this as a blueprint for packaging agent skills and system-level constraints so your agents expose modular, testable interfaces and consistent behavior (Principles 03 & 09).

Knowledge Should Not Be Gated argues for keeping knowledge human-readable in Markdown files instead of locking context into embeddings and opaque vector stores. For outcome engineering, this is a practical win: human-legible, versioned context improves inspectability, editability, and long-term maintenance of agent context (Principles 06 & 13).

Doubao and Qwen to disable humanlike, user-created agents ahead of China’s anthropomorphic AI rules reports ByteDance and Alibaba disabling humanlike and user-created agents to comply with looming regulation. Platform moderation and regulatory shifts force product and engineering changes — build feature gates, compliance workflows, and rapid rollback paths into agent deployments (Principles 15 & 10).

New AI tutor achieves 0.71–1.30 SD effect size in Dartmouth course documents large, measurable learning gains from a deployed AI tutor. This is a reminder to instrument for real-world outcomes: embed randomized evaluation, pre/post measurements, and audit trails so your agentic systems deliver and prove measurable impact (Principle 16).