Find the flaws in this computer-use agent's screenshot memory plan

Computer-use agents work by screenshotting the browser, sending the picture to a vision model, and asking what to click next. Three things in this plan look reasonable but break in practice. **Native 4K is overkill.** Every major provider has an input resolution ceiling. Anything bigger is downscaled before tokenization, so the extra pixels cost upload bandwidth and processing time without adding legibility. UI text is comfortably readable at around 1280 to 1568 pixels wide. **Append-only history overflows context.** Each screenshot costs roughly 1,000 to 2,000 tokens. Twenty screenshots eat 30,000 tokens, which is most of a 32k window before you have even reached step 20, let alone 200. The model will start dropping early turns or refuse the request. **Prompt caching is not free.** Providers like Anthropic and OpenAI offer prompt caching, but it requires a stable prefix, sometimes explicit cache breakpoints, and cache hits are billed at a reduced rate, not zero. The moment you mutate older history, the cache is invalidated. Treating cached images as free removes the only pressure to manage memory.

The plan has three independent flaws, each fatal on its own. **Native 4K wastes tokens.** Providers downscale beyond their input ceiling (Claude resizes if the long edge exceeds 1568 px; GPT-5.5 normalizes to roughly 1024 to 2048 px tiles). Sending a 3840x2160 screenshot only inflates the upload and pre-processing cost; the model never sees the extra pixels. Downscale client side to the provider's effective resolution. **Append-only history overflows the context window.** At about 1,500 tokens per screenshot, ten screenshots cost 15,000 tokens. With a 32k window, you overflow well before step 30, never mind the 200-step cap. Production agents keep a sliding window (typically the last 2 to 5 screenshots) and replace older ones with short text summaries of the observed state and the action taken. **Prompt caching is neither automatic nor free.** Anthropic requires explicit cache_control breakpoints; OpenAI caches automatically but only when the prefix is stable and at a 50 percent discount, not zero. Any mutation (evicting an old screenshot, editing system prompt) invalidates the cached prefix. Resending the same image does not magically reuse tokens. **The right pattern:** downscale at capture, keep a small recent-screenshot window, summarize older steps into text, attach cache_control to the stable system prefix, and budget image tokens explicitly.