@no_stp_on_snek: Grok 4.5 Review - I ran out of tokens for the week. Wrote an free iOS/Android app called "Hermes Go". Submitting to the…

Grok 4.5 Review - I ran out of tokens for the week. Wrote an free iOS/Android app called "Hermes Go". Submitting to the app store for review today and play store when i have the time. i ran a ~32-hour grok 4.5 session and it built me an entire production flutter app. then i dug through the local session logs to see what actually happened. the numbers are a little unreal. the output: "hermes go", a full iOS gateway client for @NousResearch's hermes agent. 33k lines of dart across 51 files. websocket realtime sync, offline-first caching, slash command and skill autocomplete, message edit/retry/interrupt, themes, haptics, voice dictation, localization in 7 languages, app store assets. from an empty directory. the scale: 104 prompts from me turned into 880 autonomous agent loops, 2,066 tool calls, and ~20k lines written across 739 edits. that is about 190 shipped lines for every prompt i typed. the part nobody talks about: it read twice as much as it wrote. 919 read and grep calls against 682 edits. instead of hallucinating the gateway API, it went and studied the hermes desktop app's source to reverse-engineer the real one. when i told it "don't act 'likely', go look at the logs", it did. my actual job in all this: product manager with a phone. i pasted screenshots and said things like "that logo looks like ass" and "the mic button looks low", and it iterated. build, push to my physical iphone, i QA, repeat. about 4 hours of model compute spread across 32 hours of me testing on device. it blew through its context window a few times to say the least, compacted, and kept going without losing the thread. long-horizon agentic coding is viable on grok now, running on my laptop, and it ships apps while i nitpick fonts. well done @SpaceXAI team! now to get apple folks to actually review it