Agent harnesses multiply while subscriptions and local models squeeze
Docker released Docker Agent (49996259). The top reaction: 'agent harnesses are turning into JS frameworks from yesteryear. All the cool kids have one.' Another asked what it has to do with Docker, and a third linked its sandbox configuration. Meanwhile an older thread (47633396) shows Anthropic ending Claude Code subscription use through OpenClaw, citing 'outsized strain.' Users replied that this is just using the capacity they paid for, and that startups get hurt. A separate thread (48542100) asks whether anyone has replaced Claude or GPT with a local model. Reports include Qwen 3.6 27B roughly matching Haiku 4.5, and others still stuck on slow tokens per second.
The pattern: harness builders keep multiplying, model vendors keep controlling access and limits, and local models are close but not a drop-in replacement. The harness is the cheap layer, and whoever owns the model controls the terms.
Sandboxing is where Docker has a natural claim, and it is the one piece of the harness stack that isn't commodity.
So what?
Don't build a business on a flat-rate consumer subscription loophole, because the vendor can close it overnight. Keep your agent layer portable across API providers and open-weight models. If you are building a harness, the defensible part is safety and sandboxing, not the loop.