Weekly findings,
without the noise.
One-paragraph briefs of material changes in agent architecture, security, governance, evidence and runtime engineering. Each brief links to the full weekly analysis and primary sources.
Harness weekly brief
Identity and runtime assurance dominated the week. NIST finalized token-protection guidance; OpenAI introduced a structured misalignment-reporting framework; Microsoft reinforced agent identity, tool governance and execution isolation; Anthropic described online and offline monitoring around a large internal agent fleet; NVIDIA’s NeMo Voice Agent exposed a new multimodal runtime surface; and FINSKILLOPS plus ActGuard strengthened promotion-gated learning and pre-execution action auditing.
Read the full weekly analysis ↗8–14 September 2026Harness weekly brief
The harness became an explicit control surface. UK DSIT published an AI Risk Management Toolkit, ASD/ACSC published dedicated Agentic AI Harnesses guidance, ENISA operationalized CRA reporting, new research showed that unnecessary tool exposure can degrade answers, Ecdysis proposed failure-pattern-driven harness evolution, and runtime projects strengthened sandboxing, OAuth, approval-state and native agent-evidence capture.
Read the full weekly analysis ↗1–7 September 2026Harness weekly brief
The first September week established the operating model for longer-lived agents: supervised bounded iteration above the worker harness, provenance-aware durable memory, meaningful human oversight rather than approval fatigue, explicit skill admission and authorization, delegated-authority graphs for OAuth identities, end-to-end authentication/PQC evidence and evaluation that separates schema validity from grounded qualified outcomes.
Read the full weekly analysis ↗