Ben's session

Ben got annoyed that Google Calendar's booking-link form makes you type in date and time ranges instead of just dragging slots on the week view, so he set an agent loose to fix it with a Chrome extension. He was testing a model he calls Luna on Max reasoning, partly because its price had just been cut 80 percent and people had been raving about it. His prompt was thin, just a rough description plus a screenshot, but it was enough to kick off the agent loop: two web searches, 55 seconds, and a mini plan came back.

He skimmed that plan instead of reading it properly, which is how he missed that the syncing logic wasn't what he wanted, the form was supposed to update automatically as you dragged, not require a manual click. The agent had computer-use and browser-use tools available and could have installed and tested the extension live, but it didn't, so when Ben installed it himself the dragging worked but the form never synced.

What followed was what he calls his frustration escalation. He typed out the issues first, still broken 14 minutes later. He tried a voice ramble with a screenshot, still broken after 61 minutes. He recorded his screen with a voiceover, pointing at the exact moments things failed, since agents can break video down frame by frame and transcribe it. The agent finally started actually testing in Chrome, but held back from touching his calendar page because it read it as his real work rather than a test target, which is exactly the page it needed to mess with.

Things got messier from there. The context window hit its roughly 250,000-token limit three separate times in one long stretch, and each compaction wiped out testing notes the agent had built up, forcing it to relearn what it already knew. Ben also noticed it had wasted time testing an old saved version of the extension, since it had been saving each revision as its own folder. He switched models to Sol on High reasoning for the coding and debugging work, fired off a mid-task follow-up message to clear up the folder confusion, and 13 minutes later the extension worked exactly as intended.

The extension does its job now, something he figures he'll use maybe twice a month. But his real point isn't about the build, it's about the steering. Reading the plan before saying build it, spelling out testing and iteration up front, telling the agent to preserve key findings before compaction, checking it's editing the right files, all the things that would have saved him an hour of looping and a pile of tokens were on him, not the model.