There is a complaint about AI systems that is repeated so often it has stopped being examined: they produce slop.
It is usually said as a verdict on the technology. Skyvern’s founder says something more precise, and the precision is the whole article: agents produce slop “because they love to do what they’re told, but they often don’t have the context required to do it well.”
Read that twice. The failure is not disobedience and it is not incapacity. It is compliance without context — a system doing exactly what was asked by someone who did not realize how much they had left unsaid.
The comparison that makes it obvious
He puts it next to hiring: “when you hire a new engineer or new employee, they try to be helpful, but they just don’t have the context for your business to really be helpful right away and they take time to onboard. Agents are the same thing.”
No competent organization gives a new hire a one-line brief and then concludes from the result that hiring does not work. There is an onboarding period, and everyone accepts it, and nobody calls the first month a failure of the candidate.
Agents get no such period. They are handed an instruction, judged on the first output, and written off. The asymmetry is not really about AI — it is that onboarding a person is a familiar cost and onboarding a system is an unfamiliar one, so the second one looks like waste while the first one looks like process.
What the context actually is
This is the part organizations underestimate, and it is why the work is architectural rather than configurational.
The context a system needs is mostly not written down anywhere. It is the reason the process has a step that looks redundant. It is which customers get an exception. It is the distinction between the two fields that appear to hold the same thing. It lives in the heads of the people doing the work, and it surfaces only when someone new gets it wrong.
Which means the work is not prompt-writing. It is finding the undocumented rules an organization runs on and making them explicit and available. That is slow, it is unglamorous, and it is the difference between a system that works and a demonstration that impressed everyone once.
What it made possible
Skyvern’s founder describes running product management, marketing, sales and customer support himself, with agents writing PRDs, managing SEO, doing content marketing, handling customer support and fixing small bugs — and says plainly it “wouldn’t have been possible without all of these agents helping me every day.”
Note the shape of that. Not one impressive system. Many ordinary ones, each doing a named job, each given enough context to do it. The result is not a replaced function — it is a person operating at a range that would otherwise have required hiring, and the range comes from the accumulation rather than from any single agent being remarkable.
What this means for an organization deciding whether AI works
- A disappointing pilot is usually evidence about the brief, not about the technology. Before concluding the tool is inadequate, establish whether it was ever told what it needed to know.
- Budget for onboarding. The same allowance a new hire gets, applied to a system, and for the same reason.
- Expect the expensive part to be extraction. Getting undocumented knowledge out of people’s heads is the work. The building is comparatively easy.
- Prefer many small systems with specific jobs over one general one with a vague remit. A vague remit is a context gap with a friendly name.
The organizations that conclude AI does not work for them have often run a fair test of the wrong thing. They tested capability. The constraint was context.
Totomoko re-architects how organizations work so that AI does the job — inside the systems they already run, on an architecture built for portability: model routing and a data layer the organization owns. Built that way, swapping a vendor is a configuration change rather than a migration.