BLOG
Anthropic lays out a four-layer agent-memory hierarchy and shows how dreaming improves recall, corrects errors, and builds navigable knowledge.
OpenAI keeps Git and rebuilds the server with Pando; Anthropic replaces the Git workflow with Scion. A detailed comparison of consistency, caching, migration, and fit.
METR's 11-week high-access review estimates Anthropic's well-specified capability growth at about 2x the historical rate, while redactions limit independent review of the strongest evidence.
With Claude writing about 90 percent of Claude Code, Anthropic runs seven layers of automated review plus risk-tiered approval, doubling merged PRs while reverts fell 40 percent. The most instructive finding is convergence: an agent author and agent reviewer once burned fifty-plus rounds on one PR, and git blame showed 59 percent of the repeatedly flagged code had been there since the first pass.
As an Amazon Associate, we earn from qualifying purchases.
