Ask HN: Are you structuring knowledge for Agent usage | Hacker News Reader

Ask HN: Are you structuring knowledge for Agent usage

1 pointtmaly3 comments

If you’re structuring knowledge for use with agents, what formats or approaches are working for you?

I want to help junior engineers on my team through a dynamic trainer.

I’m thinking about building a knowledge graph and exposing it via an MCP server so an agent can query and reason over the internal knowledge of the company.

How others are doing this? knowledge graphs? embeddings?

View on HN

What do you mean “knowledge”.

I know that sounds a dumb question but … look my take on this is one should aim for a whole-org test suite. That is can I test every input and output of a company using automated methods. And then can I create a “prod parallel” version of that company that responds the same way to the same inputs

Now of course you can’t in almost every org - but it’s possible. And once you do then the code base (that you bring up in prod parallel, and runs the company in prod) that code base represents all the knowledge of company.

Now can you do that for your part of the org ? What’s your inputs (Jiras? Emails coming in saying help? Project managers in helicopters?)

Can you capture those, replay them? What git commits did they lead to?

start with the code. Software will rip the implicit and make it explicit - its going to be a wild ride, as all those companies that implicitly cheat, will have to watch those that can explicitly represent their company in code simply execute at the speed of data centre where a nudge and a wink over racquetball will be too slow.

I never considered that idea of treating the code and company as executable knowledge.

I am trying to figure out how to structure the unstructured human knowledge like the design docs, Confluence pages, process notes into an ontology or graph that an agent or a human can query to understand how and why systems work the way they do.

Don’t think of it that way - instead think of software as the new literacy. And therefore wikis and design docs and architecture diagrams (anything that holds knowledge but it non-executable) is just illiterate scrawls. What do you expect of writing made by people who cannot read or write (code). If they could write code they would have written code. Instead they write at best Jira tickets.

So ignore it. Maybe walk over it with and LLM but then what - perhaps they are making buggy whips.

Is there a difference between a hotel that has you login on a screen and the savoy where a top hatted doorman opens the taxi door and says welcome Mr Brian?

Yes there is a difference but 99% of the difference is location and decor quality.

Maybe you don’t need the doorman if you are about to 10x the speed?

Reply on news.ycombinator.com