Josh Miller

MCP gives organizations a way to build a web that does more than serve markdown. It enables a back-and-forth — your site can expose functionality, capture intent, and participate in the agentic ecosystem rather than just feed it.

AI is going to be a multiplier of these kinds of efforts, but it will never replace community-based conversations that happen amongst agencies and practitioners. The human connections and the ability to work together — even when that means having a long view of the value of such work — are things our LLMs can articulate but would never execute autonomously.

Having gone to enough DrupalCons now to lose count, hear me when I say: the Nonprofit Summit is the single most impactful event I have ever attended during one of these conferences. See you there.

Researchers added imperceptible noise to medical images. A healthy retina becomes diabetic retinopathy. A normal chest X-ray becomes pneumothorax. The images look identical to us. The AI sees something completely different. The model found patterns that work — not necessarily the patterns we assume it found.

Wells Fargo’s CEO wanted 8 products per customer. Why 8? ‘It rhymes with great.’ Employees were rewarded for hitting the number. Result: 3.5 million fake accounts, $3 billion in fines, 5,300 employees fired. That’s the alignment problem showing up for humans without AI like we have it today.

Like a river seeking least resistance, modern AI found a different path — neural networks consume massive data and patterns emerge through backpropagation. Powerful, clearly. But it costs us something. You can’t open the hood and trace the logic anymore. This is a new kind of technical debt — one that is very, very opaque.

About 66% of our sprint work was ‘invisible’ — unpointed tickets that were diluting the signal. Once I filtered to pointed work only, a different picture emerged: our core dev team is running at 91% completion with stable velocity. The apparent decline was a data visibility problem.

Built MCP servers and clients using the Python SDK — tools, resources, and prompts. Resources and prompts were standout concepts: despite generic names, they enable powerful patterns for controlled data exposure and standardized workflows.