The Reason Most Agents Produce Slop
The founder of Skyvern reached $2 million ARR essentially solo , running marketing, sales, product, and customer support with AI agents handling most of the operational work. His company automates browser-based workflows for healthcare companies.
His explanation for why most agents produce low-quality output: they do not know anything about your company.
The agent knows what you type into the prompt. It does not know what your product does, why certain decisions were made, what your customers have been asking, or what your team tried and rejected last quarter. Without that context, it produces the same generic output it would produce for anyone.
The fix is not a better prompt. It is a better knowledge base.
The Rule That Changes How You Organize
One principle from the Skyvern approach that has broader implications: anything you do not record is not saved for the agent.
In person meetings, casual decisions, hallway conversations , none of that exists for an AI agent unless it was captured somewhere. In a remote company, decisions happen on calls and in Slack. Record the calls. Save the Slack threads. In a hybrid company, the in-person decisions that do not make it into documentation simply do not exist in the agent's world.
This pushes companies toward a specific organizational practice: recording everything that matters, in structured tools the agent can search. Not because you want an AI to read your Slack, but because the context those records provide is what makes the agent useful rather than generic.
The PRD Skill That Took Multiple Tries
The first version of the PRD-writing skill produced output the team described as slop. Good enough to recognize the format, not good enough to use.
The final version works like this: the agent starts by searching call recordings for anything related to the feature topic. Then searches Slack for communication about it. Then searches Notion documentation and customer recordings. From all of that, it drafts a first version grounded in actual evidence , not invented requirements.
Then sub-agents run adversarial review on the first draft. They leave comments. They challenge assumptions. The final step is a prioritization framework , RICE or equivalent , that eliminates the requirements that seem important but are not.
The team went from "this is slop" to "I'll actually use this to draft my features." The difference was not a better prompt on the first call. It was feeding the agent the context that a good PM actually uses when they write a spec.
The Pattern Across All Useful Agent Work
Every useful agent at Skyvern follows the same structure: it has access to the company's accumulated knowledge , calls, Slack, docs, customer data , and it produces output grounded in that knowledge rather than invented from general patterns.
The agents that do not work are the ones operating in a vacuum. Generic context produces generic output. Specific context, accumulated over the actual operation of the business, produces output that the team is willing to act on.
The implication: the value of AI agents at a company scales with the quality and organization of the company's knowledge base. A company that has been recording decisions, capturing customer interactions, and organizing institutional knowledge for three years will get dramatically more from the same agents than a company that has not.
That gap is not about the model.
It is not about the prompt.
It is about what the company already knows and whether it is in a form the agent can find.