Agent harnesses are getting swapped like text editors

The "Goodbye Claude, Hello Pi" post describes leaving Claude Code for the Pi harness, and the author says Anthropic announced AGENTS.md support a day after the decision. Commenters say the AGENTS vs CLAUDE.md split was always going to collapse. One person runs nightly local-model tests and reports that opencode beats ohmypi on code benchmark tasks. In the older Anthropic thread (47633396), subscribers were complaining that Claude Code subscriptions could no longer power third-party tools like OpenClaw. The grievances were rate limits hitting within 1:45 on afternoons and "when you fully use the capacity you paid for, that's too much."

The pattern: the model is no longer the sticky part. Instructions, skills and memory files are becoming portable text, and the harness is a thin, replaceable layer. The agent-memory thread backs this up. Commenters call memory "strictly harmful" because it fixates on the wrong things, and ask why agents that read plain text need a new spec at all.

The counterpoint is that vendors can still win on pricing and usage policy. When Anthropic limits what a subscription can do, users start comparing alternatives immediately.