The Ramp Tax
Why AI agents work for engineering and no other roles…yet.
I’ve been thinking about a disagreement between Dwarkesh Patel and Dario Amodei, and I think they’re both missing the same thing.
Dario’s position is that AI will transform every knowledge-work role the same way it’s transforming engineering. Bigger context windows, better models, more training on the work itself, and the technology will get there. Dwarkesh pushes back: coding is special. Engineers have a codebase. They have Git history, PR descriptions, tests, type systems. Every decision is documented by necessity. No other role is documented the same way. Without that backbone, the models don’t have anything unique to work with.
They’re both arguing about AI capability. The interesting question isn’t about AI. It’s about humans. A senior engineer joins a new company and ships meaningful code in weeks. A senior GTM leader at the same company takes months to ramp. Same caliber of person. Same access to tools. The difference is that one of them walks into a function with institutional memory, and the other walks into a function that runs on what lives in people’s heads.
I call that gap the Ramp Tax. It’s the cost every company pays to transfer tribal knowledge to help a new hire onboard and become productive.
Tribal knowledge is context that already exists somewhere in the building, it’s just not anywhere a new person can easily access it without “time and training”. And the Ramp Tax compounds. A B2B sales hire pays it twice (product knowledge, then customer knowledge). A new exec pays it across an entire org. And the company pays it again every time someone leaves because each person carries tribal knowledge out the door that hasn’t fully been documented for their replacement.
The Ramp Tax isn’t wisdom accumulation. It’s pattern recognition being built on top of tribal knowledge; the undocumented context that never made it into a doc because documenting it was too resource intensive. A new hire becomes valuable not because they get smarter, but because they finally have the context everyone around them has been operating on. The context was always the unlock. Us humans just carried it quietly.
Every company I’ve worked with runs on what I call Minimum Viable Documentation: the least amount of processes and SLAs a team can write down and still function. MVD is rational. Documentation is expensive, work is urgent, and humans have memory. You write down what’s strictly necessary and keep the rest in your head, because the person who you learned it from is still in the building. MVD has worked for humans because the gaps get filled in over Slack and Zoom by the author.
MVD functioned when work moved at human speed. The person who wrote the process was usually still in the building when the next person needed it, and the things that weren’t written down didn’t change fast enough to matter. AI changes the math. Work is accelerating faster than humans can document it, and the gap between what a company knows and what a company has written down is widening every quarter.
We’ve spent decades pretending this is a discipline problem. Every all-hands that ends with “we need to document better” is an all-hands that will repeat itself next quarter. My co-founder Daniel has rebuilt his end-to-end processes three times in four years, same operations, re-documented from scratch, because the old docs rotted faster than he could maintain them. My wife worked at a Series F startup where she was the only person in the org who built an end-to-end Confluence for her team, and even hers was two years out of date. Onboarding is tribal handoffs. This isn’t a failure of will. It’s a failure of design. We keep asking humans to do the one thing humans were never going to do well.
Engineering escaped this by accident. The codebase is the documentation. When Francis joined ORCA as our first full-time dev, one of the first things we did was index our repo with Devin (Cognition Lab’s AI tool) to make an AI wiki. Within a day Francis could ask the codebase anything: architecture, decision history, why a given function existed, and get a real answer. Francis paid almost no Ramp Tax. That was the moment this first clicked for me. Coding AI works because software engineering is the only function where the work itself produces the institutional memory as a byproduct. Every commit message, every PR, every test, is a documentation artifact that didn’t require anyone’s discipline to create. For every other team the memory lives in people’s heads. That’s the asymmetry the entire AI-for-knowledge-work debate is built on, and nobody is naming it.
Nobody is naming it because the people best positioned to see it aren’t looking. AI researchers are studying the models. Operators are studying the workflows. The substrate underneath, the memory the work runs on, has no owner. Engineers escaped the problem by accident, so they don’t see it as a problem. Operators feel it every day but don’t have the language for it.
The answer isn’t more documentation discipline, and it isn’t bigger context windows. It’s automated capture. The work itself has to produce the documentation, the way code does. That requires two things: conversational memory (what the agent and the human just worked through) and institutional memory (the growing repo of what the company knows). Two layers, both captured automatically as a byproduct of doing the work.
That’s what I’ve been building with Cortex. Every user gets a GitHub repo provisioned the moment they sign up. The agent writes to it automatically, capturing decisions, Q&A pairs, process docs, FAQs, competitor research, all auto-indexed and updated as work happens. The user owns the repo. If they ever leave Cortex, they walk out with the institutional memory their agent built for them. When Daniel first started using his agent without the institutional layer turned on, he told me it felt handicapped compared to mine. Every session was a fresh start. The difference wasn’t intelligence. It was memory. Once the institutional layer was on, his agent started compounding. The recursive flywheel that engineering accidentally built into codebases, that work produces memory and memory accelerates work, finally exists for a non-technical operator.
The principle underneath the mechanism is what matters more than the mechanism itself. Write-by-default, not write-on-demand. Anything that requires human discipline at the moment of capture will fail at AI speed, so the system has to write everything automatically and let the human opt out of what shouldn’t be kept, not opt in to what should. And memory has to be tiered. Some facts decay in fifteen minutes. Some compound over years. A flat memory layer either drowns the durable stuff in noise or treats every passing comment as canon. The architecture has to know the difference. These aren’t Cortex-specific design choices. They’re the rules any system trying to do this will soon converge on.
the system has to write everything automatically and let the human opt out of what shouldn’t be kept, not opt in to what should
The implication isn’t just that agents get better. It’s that the Ramp Tax stops being paid by humans. The agent absorbs the context once, on behalf of the role (product, CS, sales), and the next human who joins inherits it instantly. The First 90 Days becomes the first 9 days. Turnover stops being catastrophic. The tribal knowledge that used to walk out the door is already written down. Operational rigor stops being a function of discipline and starts being a function of infrastructure; the rigor gets installed. And the humans who used to spend their energy translating context to new hires get that energy back for the work they are truly differentiated at: aspiration, judgement, and creativity. Deciding what “good” actually means, choosing which problems are worth solving, making the moral and creative calls about what a company should be.
Dario is right that AI will reach every role. Dwarkesh is right that engineering is special. They’re both wrong about why. Engineering isn’t special because of the models. It’s special because the work writes itself down. Once that’s true for every other role, the rest of the org catches up.
The companies that win in AI won’t have better models. They’ll have better memories, the kind us humans were always expected to keep, finally written down.