OpenClaw (龙虾) was built for developers — ordinary users are paying the price

Developer design, consumer pain

OpenClaw (龙虾) was praised as a breakthrough for automating local apps. It has been reported that the same features are now draining users’ wallets and making the agent behave worse the more it runs. Why? Not because the underlying large models are dumb, but because the tool is mismatched to its new audience: what was engineered for long, stable developer workflows is being used at consumer scale without the necessary guardrails.

Three “token” black holes

Reportedly there are three structural cost drivers. First, a heartbeat keepalive syncs full-screen OCR, clipboard and long system prompts (AGENT.md and SOUL.md) every few seconds so the stateless Transformer always “knows” the machine state — great for dev reliability, disastrous for API billing. Second, OpenClaw defaults to a single, high-capability model for every task, so cheap mechanical work is handled by expensive, high‑latency models that produce verbose chains of thought and waste tokens. Third, blanket full‑screen scanning converts even a tiny UI click into thousands of image tokens because vision models tile high‑resolution screenshots into many compute blocks. The result: heavy costs and falling accuracy on routine operations.

Practical fixes and industry moves

There are straightforward mitigations for users: lower heartbeat frequency or disable it when idle; run a local lightweight model for monitoring and call cloud models only for hard reasoning; restrict scans to the active window or use pixel‑difference filters. Industry frameworks already push this “right model, right job” approach — examples include Microsoft’s AutoGen, Alibaba (阿里巴巴) Tongyi AgentScope and Baidu (百度) AgentBuilder — and several vendors are experimenting with context caching and event‑driven triggers to cut repeated uploads. So should ordinary users run OpenClaw out of the box? Probably not, unless they’re willing to tinker.

Broader context

This is also an economic story. Analysts note that tight global GPU supply and export controls have pushed cloud compute prices higher, increasing the penalty for token‑inefficient designs. Optimization is therefore both a technical necessity and a cost imperative — and until agents are re‑engineered for consumer usage patterns, many early adopters will find their crawfish getting more expensive to farm and, ironically, dumber the more they feed them.