Weekly Notes - 2026-W08

🗒️ Weekly Notes

Personal Reflection

Agents are maturing fast — new frontier models, real autonomy research, and the first serious security incidents in agent ecosystems all landed this week. The gap between "AI assistant" and "AI that acts in the world" closed a little more.

image_kw08.png

🧠 Main

  • Introducing Claude Sonnet 4.6 — Anthropic's biggest Sonnet upgrade yet brings frontier-class coding, computer use, and a 1M token context window. It's a meaningful step up from previous Sonnet releases and now competes directly with top-tier models.

  • Gemini 3.1 Pro — Google's upgraded reasoning baseline scores 77.1% on ARC-AGI-2 — double what Gemini 3 Pro achieved. A quiet but significant signal that Google's reasoning track is accelerating.

  • Coding Agents in Feb 2026 — A practical deep-dive comparing Claude Code (Opus) and Codex on context management, correctness trade-offs, and skills automation. Required reading if you're evaluating which coding agent to commit to.

  • Apple AI Wearables — Apple is accelerating three AI wearables: smart glasses targeting 2027, a camera pendant, and AI-enhanced AirPods. The physical layer of the agent ecosystem is taking shape.

  • OpenClaw → OpenAI — Peter Steinberger joins OpenAI while OpenClaw moves to a foundation and stays open source. A significant shift for the Claude tool ecosystem and a statement about where serious agent infrastructure is heading.

  • ByteDance Building US AI Team — ByteDance is hiring around 100 AI roles at its Seed division in the US, despite ongoing national-security scrutiny. The talent competition isn't slowing down.

  • Distillation Attacks on AI Models — Google was hit with 100k+ prompts in a coordinated model-cloning campaign; OpenAI is making the same accusation against DeepSeek. Model IP is now an active battleground.

  • US Military Used Claude in Venezuela Raid — Claude was deployed via Palantir in a classified DoD operation, raising pointed questions about how AI use policies are enforced when models reach the field.

🧪Research

  • Measuring AI Agent Autonomy in Practice — Anthropic analyzed millions of Claude Code sessions and found the longest autonomous runs nearly doubled to 45 minutes. Experienced users auto-approve more actions but also interrupt more — a nuanced picture of how trust actually develops with agentic tools.

  • The Problem Isn't OpenClaw, It's the Architecture — The real risk isn't any single agent framework — it's the agent + tools + marketplace combination that creates a new attack surface. Malicious ClawHub skills exposed structural vulnerabilities that will persist beyond OpenClaw itself.

  • How to Sell to Agents — When AI agents become buyers, transaction costs collapse and the rules of commerce change: machine-readable pricing, reliability scores, and per-request billing will matter more than brand recognition.

🛠️Tools

  • Manus in Your Chat — Manus brings its full agent — multi-step tasks, voice, and file handling — directly into Telegram. No setup beyond a QR-code scan, which removes most of the friction for non-technical users.

  • VisionClaw — An open-source iOS/Android app that gives Meta Ray-Ban glasses real-time Gemini voice and vision AI, with optional OpenClaw tool execution. The simplest path yet to a capable AI wearable using hardware you already own.

  • Qwen3.5 — Alibaba releases Qwen3.5-397B-A17B: a natively multimodal MoE model with 1M context, 201-language support, and benchmark scores that challenge GPT-5.2 and Gemini 3 Pro. Open-weight competition at the frontier is real.

  • xAI Grok Build Parallel Agents — Grok Build gains support for up to 8 simultaneous parallel agents and an arena-mode evaluation layer for comparing outputs. Worth watching as a template for how multi-agent UX might evolve.

🌅Closing Reflection

A week where "agents" stopped being a buzzword and started showing up in DoD operations, active supply-chain attacks, and billion-dollar infrastructure deals. The two pieces worth revisiting: Anthropic's autonomy research for the empirical view of where agent behavior actually is today, and the architecture-level security essay for where the risks are heading.

🙏Thanks & Contact

Thanks for reading! If you have suggestions or feedback, I'd love to hear from you via my contact form. See you next week!