Measuring AI agent autonomy in practice — Anthropic publishes metrics and findings on how users grant and manage agent autonomy across real deployments. This matters because it surfaces concrete autonomy patterns and domain-specific risks you must monitor post-deployment, giving you a starting signal for governance and validation (Principle 16).

New Research Shows AI Agents Are Running Wild Online, With Few Guardrails in Place — MIT CSAIL documents widespread agent deployments with minimal safety frameworks, scarce disclosure, and high autonomous risk in browser agents. This matters because it exposes the operational failure modes you need to defend against—disclosure, operator accountability, and runtime guardrails are immediate engineering priorities (Principle 15).

9 Observations from Building with AI Agents — Tomasz Tunguz distills practical rules for reliable agent systems: pick top models, version prompts, centralize context, and automate closed-loop improvements. This matters because these tactics map directly to day‑to‑day outcome engineering work—traceability, repeatability, and continuous improvement make agents dependable in production (Principles 06 & 13).

Gemini 3.1 Pro: A smarter model for your most complex tasks — DeepMind rolls out Gemini 3.1 Pro with big jumps in complex, multi-step reasoning and broad availability across API and platform surfaces. This matters because stronger reasoning and larger context windows let agents own longer workflows and reduce orchestration brittleness, changing how you partition tasks between agents and humans (Principle 09).

Consistency Diffusion Language Models: Up to 14x Faster Inference Without Quality Loss — Together AI shows CDLM techniques that cut inference latency up to ~14x via trajectory distillation and block-wise KV caching. This matters because dramatic inference and cost improvements unlock higher-frequency, low-latency agent loops and make long-running, context-heavy agents economically viable at scale (Principle 07).