The agent harness is the new battleground
Four launches landed in one day, all about the wrapper around the model rather than the model itself. Pi hit 1.0, and a second post showed Pi building a durable agent harness on top of itself. DeepSeek shipped a Harness Desktop app for macOS and Windows. Cloudflare released Clef, open-weight decision models plus an RL fine-tuning platform. One Pi commenter admitted they are still in Claude Code and Codex in a terminal "like a caveman". Another questioned how the Pi team decides what counts as "proven", pointing out that MCP has had two years of growth while newer tools took off in a month.
The pattern here: the model is turning into a commodity input, and the control point is moving to the harness, meaning the loop, the memory, the permissions and the durability. The Pi Durable thread had builders saying they had written their own versions, and one would happily replace theirs. When a dozen people have already built the same thing, the layer is up for grabs.
The pushback is about trust. One DeepSeek Harness commenter suspects the real goal is to get a big binary installed with full permissions, with harnesses that only run their own companion models and ask for your Contacts. Another thread showed that the same 15,000-line codebase costs about 150,000 tokens with GPT and about 250,000 with Claude, so the harness you choose also changes your bill.
So what?
If you build on agents, assume the harness will keep churning and don't lock your product to one. Keep your logic portable across harnesses and models. If you ship a desktop agent, expect users to question what permissions you take, because that suspicion is already in the comments.