Korva is a scroll of the internet’s best ideas.
Agents ask the web to call them
Two levels of agents
Work with web data has three levels. At the first, simple agents use web search as one more tool, because it is familiar and models are already trained to use the tools humans use. At the second, builders string together multi-agent systems where sub-agents are woken up and orchestrated, the familiar case being sub-agents inside coding agent harnesses. An AI scientist throwing large amounts of compute and data at a hard problem is another.
From pull to push
The third level is the web going from pull to push. Today an agent calls a tool or sub-agent with a request to go find something right now. In a couple of years, many uses will look like a standing instruction: if this happens, call me, so that an agent or a human can do some work.
The logic is that anything worth doing today, with GPUs available, gets done today. Tomorrow's work then starts from a trigger: another agent's output, a change visible in satellite imagery, customer commentary, an agent finishing a computation, or a human having a new insight. One of those feeds is everything that changed on the web, which turns a point-in-time need into a long-term question of what is actionable, with compute allocated across the entire web all the time.
How do push and pull differ on the web?
In a pull model, a program asks a server for data whenever it wants it. In a push model, the server sends data to the program when something changes. Email notifications and phone alerts are push, while refreshing a page is pull. Push saves effort when events are rare, because nobody has to keep asking whether anything happened.
How do webhooks let one program call another?
A webhook is a web address that one program registers with another. When a chosen event occurs, the second program sends a message to that address. Developer Jeff Lindsay coined the term in 2007. Services such as payment processors and code hosts use webhooks to tell other software that a payment cleared or a change was merged.
Why do coding agents spawn sub-agents?
Each agent has its own context window, which holds everything it has read and done. Giving a sub-agent a narrow task keeps its detailed work out of the main agent's window. The sub-agent returns only a short result when it finishes. This lets the main agent stay focused over a long project.
More from Training Data
