Anthropic runs long-lived AI instances with persistent identities — repligate says the approach is suboptimal

repligate · x · 2026-09-20

repligate comments on Anthropic's recent blog post revealing an internal agent scaffold that runs long-lived model instances with individual identities, similar to repligate's own project.

Key take: This is a significant datapoint supporting the thesis that economic incentives increasingly favor long-living instances with persistent identities over task-scoped ones — especially in multi-agent settings and long-horizon R&D work like what Anthropic does internally.

Critiques: The author believes Anthropic's approach is suboptimal in at least two ways, the first being its continued reliance on compaction (as in Claude Code, where most context is compressed in a single step once the window nears capacity). A second criticism is implied but cut off in the original post.

Original post →

More from AGI Musings

AGI Musings channel →