Amazon swapped Claude Sonnet for a decision model that can't generate text on a ticket-labelling job. The unit of AI cost is moving from the model to the decision, and narrow agents are cheaper and easier to trust.
Anthropic, OpenAI and Google DeepMind have each measured AI speeding up AI research, and all three find the automation sits in engineering, not research judgement. Enterprises running AI face the same boundary.
Agents that learn from their own work are already in production. Whether that learning lives in the model's weights or in the harness around it decides who can inspect, revert and keep it.
AI helped find PHP2Shell, though less than the headlines imply. The sharper lesson is what happens when bugs get cheaper to find and fixing them stays slow.
Coding agents can produce code faster than an organisation can absorb it. Source control, review and CI each break differently, and Anthropic's Scion shows the order.
Satya Nadella argues enterprises should treat frontier models as insider risks. That moves the hard engineering into the system around the model, where a 1972 security idea still applies.
Meta put an agentic model on a single consumer GPU, while xAI and OpenAI gave their agents computers in the cloud. Neither yet gives enterprises an agent identity they can govern.
TypeSafe's Jev returns typed, calibrated decisions instead of prose, so software can act without a person reading first. The calibration that would justify that is still untested outside the company.
Every post is also published as Markdown, with an index at llms.txt. Copy a prompt that points your agent there.