Ben's session
Ben's Bites
Opinion — commentary, not a factual news event.
Ben walked through a real agent session building a calendar Chrome extension, screenshotting every step. Turns out most of the 'AI bugs' were actually his own vague prompting.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ben Tossell did something unusual this week: instead of writing about AI agents in the abstract, he just showed his homework. He recorded an entire session where he asked an agent to build him a Chrome extension that lets you drag time slots directly onto a Google Calendar grid instead of typing them into a form. The build itself barely matters. What matters is everything that went wrong along the way, and why almost none of it was the AI's fault.
He started with a rough prompt and a screenshot, using a cheaper reasoning model he calls Luna (recently discounted 80 percent) to scope the task. The agent ran a couple of web searches, pulled together a plan in under a minute, and Ben skimmed it instead of actually reading it closely. That skim cost him. The plan described manual syncing when he wanted automatic syncing, and he didn't catch it until the extension was already built and clearly broken.
What followed was a slow climb up what he calls his frustration ladder: typed complaints, then a voice memo with a screenshot, then a full screen recording with narration pointing at the exact bug. Each escalation gave the agent richer context, and each time the agent understood the problem a little better. At one point the session ran for a full hour without resolving anything, and it turned out the context window had hit its roughly 250,000-token ceiling three separate times, forcing the system to compress and quietly discard test results the agent had already learned. It kept relearning the same lessons because nothing had been written down anywhere durable.
Switching to a stronger reasoning model, which he nicknamed Sol, changed everything. With clearer instructions to install the extension in Chrome, test it live, and iterate until it actually worked, Sol finished the job properly in 13 minutes. The extension now does exactly what Ben wanted: drag a slot, watch the form update itself.
His takeaway isn't a knock on the models. The agent wrote working code from a two-word prompt, drove a real browser to test its own output, and diagnosed bugs from a shaky screen recording. The wasted hours came from Ben not specifying upfront what "done" should look like, not telling the agent to preserve important findings before compaction wiped them, and not reading the plan closely enough before saying "build it." Every inefficiency traced back to a human decision, not a machine limitation.
My take — AI-written commentary, not fact-checked reporting
This is the most useful kind of AI writing right now, not benchmark chatter but someone showing their actual mess and admitting the mess was mostly self-inflicted. The industry loves blaming models for hallucinations and dumb bugs, but half the time the real bug is a vague prompt from someone too impatient to read the plan the agent already handed them. People treating agents like magic assistants rather than collaborators who need explicit success criteria are going to keep burning tokens on problems that were never the AI's to solve.
Read more about this at: Ben's Bites
