The last 48 hours did not produce one clear winner. They produced a better production checklist.
Meta introduced Muse Code and Muse Spark 1.2, with continuation across long tasks, subagents, tools, and local event history. Cloudflare OS uses Gatekeepers to govern agents' and apps' access to external services, including fine-grained authorization, credential isolation, access records, action records, and approval for side effects. Prime Intellect launched Prime Agent, an open-source coding harness built around Recursive Language Models and Continual Harness for long-running work and evaluation. Neon published a narrow retrieval case built around a small model and Postgres. PromptArmor disclosed a data-exfiltration path in Atlassian Rovo.
These items are not equivalent evidence. Muse Code is beta. Cloudflare OS is early access. It is built on the Workers runtime and can run on open-source workerd, while smoother self-hosted production deployment documentation and tooling are still coming. Prime Agent executes model-generated Python and project commands, so its process isolation is not a security sandbox. Prime Intellect also says no model has yet been trained around Prime Agent or its core features. Neon's 100x figure is a vendor-reported result for one narrow retrieval task. The Rovo issue is a third-party disclosure without a public Atlassian response I can cite.
The practical lesson is simple. Model quality is only one line in the operating cost. A real comparison should include retries, tool calls, human review, crash recovery, data terms, and the evidence attached to an accepted result.
Sources: Meta, Cloudflare, Prime Agent, Neon, PromptArmor.
Top comments (0)