Nous Research's Hermes Agent has officially added a macOS computer control feature.
In practice, it directly rivals OpenAI Codex's "background control." Under the hood, it integrates the open-source driver cua-driver (which we've covered before), which reverse-engineers Apple's private API to send commands directly to target processes. That means when the agent is searching your email or writing code in the background, your actual mouse won't jump around and the current window won't lose focus — human and machine can work side by side on the same computer without stepping on each other.
Because computer control relies heavily on continuous screenshots, the token bill can balloon fast. Hermes tackles this with a four-layer context compression mechanism at the framework level: it aggressively strips redundant frames, keeps only the last 3 screenshots in memory, and clears old cache server-side. According to official estimates, running 20 consecutive steps at 1568×900 resolution drops context consumption from about 600,000 tokens to roughly 30,000 tokens — a 95% reduction.