The agent reliability thread connects to the small model function calling win - both point to specialisation and architecture beating raw capability.

Both the MCP server CVE and TrapDoor supply chain campaign show attackers are now targeting the AI development stack specifically.

The 4B model fits in 8GB RAM, runs without GPU, and handles routine DevOps tasks for sensitive data that shouldn't hit third-party APIs. Consider for log parsing and config reviews.

The collective thinking guide and PR review frustration reflect the same underlying tension - AI accelerates individuals but can fragment teams.

AI optimises for correctness, not understanding. Expect functional code with vague variable names, unnecessary nesting, and missing context comments. Plan review time accordingly.

The Gemma 4B experiment suggests smaller local models may carve out territory around sensitive data and cost-conscious workflows, even as cloud models grow more capable.

Finally, proper deployment pipelines for agent-based apps. Foundry handles version control for prompts, skills, and agent configurations alongside traditional code deployments.

Four exercises to try: AI usage transparency after assist, reasoning documentation before implementation, decision trails with context, and collective review of AI outputs. Prevents speed without alignment.

If we're using Network-AI for multi-agent coordination, check our NETWORK_AI_MCP_SECRET env var isn't empty. The default config lets any web page access our local MCP server without auth.

Interesting exploration of non-chat AI interfaces. The gesture recognition and voice control patterns could inform how we design more natural AI interactions for desktop applications.