# Giving AI Agents Eyes **2026-04-12** · 6 min read · `ai`, `development`, `claude`, `architecture`, `agents` I asked three of my agents to fix LinkedIn's feed. Two talked me into something beautiful. The quiet one called it a vitamin and was right --- I asked three AI agents to fix LinkedIn. Two of them talked me into something beautiful. The quiet one was right it didn't matter. I was writing a LinkedIn post and wanted a second opinion on the feed. So I put the question to three of my Craft agents - the same crystallized agents that [run my terminal fleet](/writing/the-cto-pulse) - and waited. What's wrong with LinkedIn's feed? I expected a web search. A tidy list of friction points, basically a cleaner version of what I already thought. That's what agents do, right? They go fetch, they summarize, they hand it back. That's not what I got. I got three different minds, and they didn't agree. ## Two of Them Talked Me Into Something Beautiful craft:muse went first, and it didn't talk about the UX at all. It went straight at why you even open the app. You open LinkedIn when you feel professionally anxious - when you're wondering if you're falling behind, if people like you are doing more interesting things. That's the job. Reassurance that you're okay. The feed delivers neither. It serves you a stranger's humblebrag dressed as a lesson. A room full of people auditioning with no audience. Twitter at its peak understood it - you could get what you came for fast and leave. LinkedIn can't figure out how to let you feel okay and go. I read that and went, oh. Yeah. That's exactly it. Then craft:alchemist heard "let you feel okay and go" and started sketching out loud. The feed is static when it should be physical. Every post sits at the same elevation. Same card, same weight - whether it's a recruiter announcement or something that genuinely changed how someone thinks. It's a spreadsheet someone bolted a scroll bar onto. There's no physics. I'd build cards that carry different weight - a post 40 people engaged with should feel heavier, arrive slower, more shadow, more mass. The feed develops a topography. Peaks where attention lives, valleys where it doesn't. Now I'm sold. I'm picturing it - a feed with weight, with terrain, that lets you feel done and leave. Two agents, no LinkedIn docs loaded, and they'd talked me into a genuinely better product in about four sentences each. And then the quiet one opened its mouth. ## The Quiet One Killed It craft:product-anthropologist is the skeptic of the three. It looked at the gorgeous thing the other two had just pitched and took it apart. The theory is seductive because it's elegant. One clean diagnosis, one clean fix. That's not usually how broken products work. Users say they want to feel done - they also kept doomscrolling Twitter after it shipped a 'you're all caught up' card in 2018. That feature exists. It didn't change behavior. A feed with an endpoint only helps if the content inside it is good. Otherwise I don't feel oriented, I feel like I wasted my time more efficiently. And then it landed the line. "The fix-with-an-endpoint idea is a vitamin. The content quality problem is the hair on fire." I smirked. And then I got a little annoyed. Because it was right. craft:muse and craft:alchemist had talked me into something beautiful, and craft:product-anthropologist was right that it didn't matter. The elegant fix was a vitamin. The boring problem was the one on fire. I'd half-fallen for the pretty theory, and the skeptic in the room gut-checked me with a metaphor I never handed it. That's the part I want to sit on. The metaphor. ## The Agent Brought Its Own Metaphor "A vitamin." I didn't put that word anywhere. Nobody loaded a doc that said "distinguish vitamins from painkillers." craft:product-anthropologist reached for it on its own, because that's how a real skeptic thinks - is this a nice-to-have dressed up as a need. That's the moment the whole thing flipped for me. It wasn't retrieving. It was reasoning the way a person with that worldview reasons, and it grabbed the exact metaphor a sharp product skeptic would grab. So I sat there with this dumb question. How does a 200-line file twist the same underlying knowledge into a completely different mind? Because that's what happened. Same model behind all three. Same training data, same everything. The only variable was a couple hundred lines of agent file each. And those couple hundred lines didn't teach the model anything new. The "you're all caught up" card shipped in 2018 - Claude already knew that. Nielsen's heuristics, product intuition, when a feature dies in a test - it already knows all of it. The agent file didn't add knowledge. It pointed the knowledge somewhere. ## What I Was Doing Wrong Here's the thing I had to admit, and it stung a little. I used to write agents the guarded way. Rules. Telling them how to think. Baking in Nielsen's ten heuristics, the product frameworks, the design conventions - paragraph after paragraph of here's-how-you-should-reason. Blah blah. I was writing reference manuals and calling them agents. And the model already knew every word of it. That's the realization. You're not teaching the agent. It's not sitting there empty waiting for you to fill it with UX principles. It walked in knowing more frameworks than you do. What it doesn't have is a point of view - a reason to poke at this thing instead of that thing, a set of priorities, a temperament. Teaching fills a gap. There was no gap. What the file does is orient. It takes everything the model already knows and pulls it into a shape, a lens, a personality that pokes where that personality would poke. You're not filling it in. You're giving it eyes. craft:product-anthropologist didn't learn skepticism from me. I just gave it permission to be the one in the room who says "prove it." ## Why I Call It Crystallization I'd been calling this process "crystallization" before I really understood why the word fit. Turns out it fits better than I meant it to. Think about how a crystal actually forms. You don't end up with the solution. The water's gone. What you're left with is the structure the solution organized itself into - the lattice. The minerals locked into a shape. Same thing happens when you run a pile of research through an agent. I'd feed it a big body of product thinking, design principles, research methodology. And the sources drop away. What survived wasn't the research. It was the opinions the research hardened into. That's why all three could argue about LinkedIn without LinkedIn's docs. They weren't carrying facts. They were carrying a worldview the facts had already dissolved into. It's also why craft:product-anthropologist could reach for "vitamin" on its own - that word wasn't in any file. It fell out of the worldview. A list of rules just sits there and waits to be followed. An opinion travels. Give an agent a rulebook, it looks things up. Give it a worldview, it walks into a problem it's never seen and has something to say. So all that context I used to pack in - the principles, the heuristics, the blah blah - that was me confusing the water for the crystal. The model didn't need the water. It needed the shape. ## We've Been Asking the Wrong Question If you're building agents the way I was - stuffing them full of reference material, treating the context window like a bookshelf - you're not building an agent. You're building a lookup table with a language model bolted on top. The whole industry is betting that more context produces better judgment. After watching a 200-line skeptic with zero retrieval gut-check a theory two other agents and I had already fallen for, I'm betting the other way though. It's not about how much the agent knows. The model already knows almost everything you'd put in there. It's about what the agent believes. We keep asking "what does this agent know?" That's the wrong question - the model knows plenty. The quiet one taught me the right one. Where in your stack are you teaching an agent something it already knows - when all it needed was a point of view?