... question: given this input, did the model produce the right final answer? Your product almost never lives or dies on that. It lives on the path. An agent that lands on the correct answer by calling the wrong tool, backtracking twice, leaking its reasoning into the user's face, and taking nine seconds to do a one-second job scores as a pass and ships as a failure. The eval that actually correlates with "does this work" is the trajectory ...

... confused user can abandon the product: the first-run promise, the decision points, and the recovery states. Everything else can begin with a consistent system and improve through observation. Record a few first sessions and look for hesitation, backtracking, or abandonment; those patterns turn make it better into a concrete design brief. That gives both AI output and a designer something real to work from. Comment from Samran Elahi(@samran_elahi): I lean on AI for the base ... ... Where I ve landed: brand foundation is a front-loaded cost, not an ongoing one. A few focused hours defining the mark, palette, and type system once, then months of not having to think about it again versus

... product mentioned in ChatGPT or Perplexity answers, and the data keeps pointing somewhere that makes me question the whole "GEO/AEO" hype. Ahrefs looked at 75k brands and found branded mentions correlated with AI visibility way more than backlinks 0.664 vs 0.218. It's correlation, not causation, so take it with salt. But the gap is big enough that I can't ignore it. Which makes me wonder if we're overcomplicating this. Maybe AI visibility isn't a new discipline ... ... move is to spot the anchor source and focus on creating fresh, highly-citeable counter-content targeted specifically at that exact sub-topic to win the retrieval context. Comment from Saied Alimoradi(@saied_alimoradi): @voiqu   ok the perception backtracking

... next week or for Agent that should see significant improvements by default, but some additional context that might be helpful to you: 1. Why it happens: The Agent (language models generally) is worse at fixing something it broke vs doing it correctly on the first shot. When you ask the agent to do something, it doesn't, then you follow up with a "it's still not working" or a "fix it", or something, unless there's explicit console logs ... ... write more code which tends to compound the problem. 2. Things that'll address it, all coming soon: better debugging and reasoning steps, better identifying a loop state behavior and bringing in a reasoning or stronger model to debug, backtracking