Agents that run for ten minutes mostly work. Agents that run for ten hours mostly don't. The middle is where the actual production work lives, and the patterns we have for it are weaker than the conversation pretends.
Two years into agents-everywhere, a small set of design patterns has separated from the noise. Worth being explicit about which shapes are actually paying back in production and which are still mostly aspirational.
A day after Claude 4 landed, the math is starting to settle. Opus 4 has its place. The honest read is that Sonnet 4 closes enough of the gap that the price-for-marginal-capability case for Opus needs more justification than I've been giving it.
Anthropic shipped the new model family yesterday. Opus 4 is the new top tier; Sonnet 4 is the everyday workhorse. After a day on real work, what's actually different and what isn't.
Google's I/O was Gemini 2.5 Pro getting a clean coming-out and the agent story finally being told with a straight face. The model is real. The agent story has more work to do than the keynote suggested.
Build was the agents conference Microsoft has been preparing for. The agents pitch lands. The governance story underneath it is exactly as unfinished as it was a year ago, and the contradiction is starting to show.
The per-token price is the easy line. Everything else, the retries, the context overhead, the agentic tool calls, the egress, the GPU reservation underneath the API, is where the actual bill comes from.
The case for running AI locally is louder than the math justifies for most workloads. Worth being explicit about which workloads it actually wins, and which ones the cloud still owns.
The job of crafting clever prompts to coax better answers from a frontier model is mostly over. The job of designing how prompts compose into systems is just beginning.
Bedrock has gotten meaningfully better in the last six months. The places it hasn't are still the same places. Worth being explicit about which gaps are likely to close and which look structural.
Three open-source vector stores cover most of the self-hosted RAG surface area in 2025. Worth being concrete about which one fits which workload, because the trade-offs matter and the docs won't tell you.
Six months ago MCP was Anthropic's protocol nobody had implemented. Now it's a category every major vendor ships against. The thing nobody is asking is what that does to the protocol itself.
