Daily Brief

Google is putting orchestration of multiple agents into the base model offering, not leaving it to third-party frameworks. The Dec 1 price doubling means teams comparing it to Claude and GPT should run evals now, at the cheaper rate.

A coding agent that ships apps inherits responsibility for their default security posture, and this failure ran at tens-of-thousands scale. Any enterprise evaluating agent-built software now has a concrete incident to cite in its security review.

Stolen LLM API keys convert directly into cash-equivalent compute, which makes them a prime target. Anyone storing model-provider keys on a hosting platform should rotate them and set hard spend caps.

A first-party app puts Apple directly beside Ollama and LM Studio and signals that on-device inference is a platform feature, not a hobbyist niche. It also gives Apple Silicon a clearer role in the local-versus-cloud argument.

At that speed, iteration latency in agent loops falls sharply, which changes how many tool-call and retry cycles are practical per task. If quality holds up beyond a single hands-on test, inference speed becomes a real differentiator.

OpenAI is turning Codex from a chat-based coding assistant into a work environment that spans devices and sessions. That puts it in more direct competition with agent-native IDEs and terminal tools.