I have been working on search systems for a while, and I always wondered: What if search, instead of matching documents, matched workflows and ran those workflows?
I called it “Search that executes” — as opposed to search that retrieves information.
When AI agents showed up a few years ago, I realized that those workflows could be agents — capable of understanding a user’s query and taking branches dynamically. That’s when I got excited enough about the idea to implement it.
To make this work, it needs more than just search: a structured framework for defining agents (so the system “understands”), a controlled runtime for executing them with guardrails, and an end-to-end system with a UI to request and receive user approval before taking critical actions.
It made sense that all this plumbing should be handled behind the scenes, letting programmers concentrate on the agent’s business logic. Many design decisions followed directly from that.
Agents are created in a JSON structure. There are 23 “commands” that define the units of work an agent performs. Python expressions can be used in strings within curly braces.
Because the whole thing has a structure/schema, AI generates valid agents from English descriptions: “Create a RAG agent to call the company purchase order system”, “Create a tool-calling agent to …”. You can check this out in the live demo on the website.
The developer creates an agent, describes it in plain English, and publishes it in the system. Users just submit a request. Search matches the request to an agent, and the system runs the agent. The demo also shows searching for and executing agents.
That’s the central idea. The website has more information on other aspects. Despite the enterprise-y look and feel of the website, the system was really created by just me over the last couple of years. I would love to answer any questions and hear your comments and feedback.
I called it “Search that executes” — as opposed to search that retrieves information.
When AI agents showed up a few years ago, I realized that those workflows could be agents — capable of understanding a user’s query and taking branches dynamically. That’s when I got excited enough about the idea to implement it.
To make this work, it needs more than just search: a structured framework for defining agents (so the system “understands”), a controlled runtime for executing them with guardrails, and an end-to-end system with a UI to request and receive user approval before taking critical actions.
It made sense that all this plumbing should be handled behind the scenes, letting programmers concentrate on the agent’s business logic. Many design decisions followed directly from that.
Agents are created in a JSON structure. There are 23 “commands” that define the units of work an agent performs. Python expressions can be used in strings within curly braces.
Because the whole thing has a structure/schema, AI generates valid agents from English descriptions: “Create a RAG agent to call the company purchase order system”, “Create a tool-calling agent to …”. You can check this out in the live demo on the website.
The developer creates an agent, describes it in plain English, and publishes it in the system. Users just submit a request. Search matches the request to an agent, and the system runs the agent. The demo also shows searching for and executing agents.
That’s the central idea. The website has more information on other aspects. Despite the enterprise-y look and feel of the website, the system was really created by just me over the last couple of years. I would love to answer any questions and hear your comments and feedback.