jhgaylor

My agents need three verbs to help me write code: plan, implement, verify. Planning happens locally with the model and my issue tracker. Implementing happens across a fleet of coding agents. Verification has been the painful one. I ended up using GitHub Actions as my trusted environment for testing untrusted code, which works but is miserable ergonomics for an agent: batch execution, a minutes-long feedback loop, and no way to poke at a failure interactively. Unfortunately, verification is the step where the agent needs to iterate quickly.

To me, the interesting part isn't the isolation, RuntimeClass has given us gVisor/Kata pods for years. It makes it easy to deploy this new type of workload where an agent claims a pre-warmed sandbox, gets a prod-like environment with a stable identity, it can hibernate when idle and be thrown away when it has served its purpose.

Replacing CI for the verify step doesn't seem to be the marquee use case, but I'm going to set it up and try it.

This is at the bottom of section 8

> The Large Language Model MUST NOT have access to an agent's credentials or to credentials that may be needed to access tools and services. This prevents the Large Language Model from using, exposing, or being manipulated via prompt injection into disclosing the credentials.

I knew this but I just went ahead doing it wrong anyway. As a stop gap I gave all my secrets handles so the LLM and I could discuss them but I am just pretending that my agent is trustworthy. We have some big names here so hopefully there is better tooling in our future.

I am not a particularly good writer, so I started using llms to help me write. At first that was by having the llm convert my rough draft into something ready for editing. As I edited a bunch of this work I found it still wasn’t something I wanted to put my name on.

Now I use llms to help me research, outline, refine vocabulary and grade my work but also have stopped letting it speak for me. I’m much happier with the result.

MCP is a new protocol from anthropic to standardize sharing tools and context with LLMs. Before, the tool calling api from openai was standard but tool makers all built their own mechanisms for defining and sharing tools.

It's a bit of a stretch but MCP is to LLM enabled applications what REST is to web applications.

What was happening before was if you built tools using langchain, you'd have to rewrite them for crewai, cursor, etc. Now, we have a way to share tools, resources, and prompts with applications built using different frameworks.

I was trying to find a way to thank xpe more privately but this is evidence I should just go ahead and do it. So, thank you too.

Thanks xpe, I appreciate you jumping in here. I was struggling to find the words here and I think you did a wonderful job both championing the intent of the post as well as articulating why I found it difficult to engage. You've given me tools to use going forward.

I ended up building the first couple of iterations of this tool just to stop entering the same information into Claude for every new conversation.

By connecting an assistant to a job searching api, a database, and context about myself I am able to create a prompt such as "find interesting jobs for jake. maybe something in the ai space?" and in a few minutes I can browse a curated list of potential job matches.

By connecting the assistant to text to speech and speech to text tools and context about myself I can provide a the job description in my prompt and request the assistant play the role of an interviewer. This has been much nicer than practicing in the mirror.

I think that for the next few weeks/months that a hiring team connecting to my mcp server will play out well for me but I think you're in the right ball park. It will be because I was able to show that I can extract value from technology.

> the `candidate-info://website-text` has a bit of marketing puffery like we don't usually see on resumes. I'm wondering whether that's intended to influence the AI tool behavior.

I actually wrote the marketing for the humans. That site predates this ai native resume. My thinking is that by putting a little sell into my site I can show off another aspect of my skillset. I used to have a standard bio site with a portfolio but it was a wall of text and needed a refresher.

> As a simpler solution

llms.txt seems to work pretty well. I am sure there are ways to increase the quality of an llms.txt but I started by simply joining all the text data I already had together and asking an llm to make an llms.txt out of it. From there I've been "manually" editing it. Often with Claude's help.

> It could be under a `/.well-known/` URL

I am hoping we start to see a lot more use of this. We already have a pretty good set of tools to do discovery so let's use them.

That's what I mean but I wouldn't represent it as being me the human speaking. We can just upgrade from text to text to speech to speech (or any mixture) while still using the LLM. And for style, I can use my voice instead of Microsoft Sam.

I think if you write the first blog post about this you get to name the law.