The jagged frontier concept is key - models excel at latest research whilst struggling with basic tasks. We need to test our specific use cases rather than assuming general capability.

Edge deployment getting more efficient. If we're running inference on mobile or embedded hardware, these Hexagon optimisations might actually be noticeable.

Anthropic's finance agent templates aren't just demos - they're production cookbooks for regulated industries. The progression from general-purpose models to domain-specific agent architectures suggests the market's moving beyond 'AI can do anything' to 'AI can do specific things really well'.

Smart security model that actually helps us sleep better. Separating production credentials from dev environments reduces the blast radius when someone inevitably commits an API key.

Signals Google's serious push into creative tooling beyond just APIs. The production support for winners suggests they're building a complete creative pipeline, not just individual tools.

Refreshing honesty about AI complexity from someone who actually builds systems. The progression from extract → interpret → personalise shows a solid framework for evaluating AI capabilities.

DroidSpeak's KV cache sharing could dramatically reduce inference costs for similar models. Worth tracking if we're running multiple fine-tuned variants of the same base model.

Simon's right - it's not ethical to waste real humans' time with AI experiments they didn't consent to. We need to be careful about where our AI systems interact with the outside world.

This changes how we approach document search with charts and diagrams. Instead of OCR preprocessing, we can index visual content directly and get proper citations linking back to source images.

Solid reference architecture for enterprise RAG deployment. The progression from manual processes to AI-driven analysis shows realistic implementation patterns we can adapt.

The Craft Agents desktop workspace signals an interesting divergence - whilst everyone's building cloud-first agent platforms, there's real demand for local-first agent tooling. The MCP integration means we get the best of both worlds: local control with cloud service connectivity.

The title's right - we're moving from prompt crafting to agent architecture. The ADK patterns show how to structure skills and tools for systems that actually scale beyond demos.