I'm trying to build a small, hobby agentic system locally, looks like it is an order of magnitude more difficult to create your own full secure agent than I had previously thought.

For example, they can't tell you in one sentence what they replaced the "coding ability" screening with.

More enshitification.

I understand the AI-aided productivity gains but don't take people for fools. Running AI models in a sandbox environment takes technical know-how. Companies never hired people just for coding ability.

I hadn't thought about compounding at the time of writing this essay. One of the reasons was that I just felt like grabbing people by their necks and thrust their face in the screen while shouting "see...see..you can just do things.. don't wait.."

Sometimes, it is better to take the immediate, most feasible action: doing the thing in front of us. Then doing the next thing. Then the next..

Comments URL: https://news.ycombinator.com/item?id=47975316

Points: 3

# Comments: 2

https://gist.github.com/bhagyeshsp/510f66ae3d8eb3f826dfd8273...

Modify as you find fit.

Just read your text here, now that makes sense. It is made for your taste, solving a particular problem. If it is working for you, great! If you want more people to adopt it, it is a different discussion altogether.

I didn't get the "miles of DNA" reference. A single strand of DNA is approx 3 meter in length when uncoiled. Now I'm thinking how many strands may be replicated at a time.

But of course, for that the developer will need to make each command open in a modal instead.

I was creating an AI chat companion for one of my products and this is the compilation of my decisions and reflections. Plenty of technical parts that you'd like to look into.

*Things I think worth highlighting*

1. Cloudflare Workers

2. Custom static site for interface

3. Full system prompt at the beginning: 17000 tokens -> Ultimately 2500 tokens

4. Tried two LLMs: one as a router LLM to inject selective context + another as the main model

5. Router was a bad idea. Main issue: 6000 ms latency

6. Product matched Embedding approach perfectly (I didn't know at the time, learnt it)

7. Context distillation was a huge learning: remove all semantic words from the prompt

8. Used Promptfoo for benchmarking

9. Cosine similarity score has to be understood for your own data, don't take any random score from the internet (internet and AI suggested 0.7 similarity score, mine turned out around 0.2-0.25)

10. Found OpenAI-4o-mini the best conversation model for my case

---

Questions? Any fellow travellers having gone through the same pain?

Comments URL: https://news.ycombinator.com/item?id=47963016

Points: 1

# Comments: 1

This is the main hook in AGENT.md file:

   ## 3. Initialization Protocol (Read This First)

   Upon being activated in this directory, you MUST:

   1. **Read `CURRENT.md`** to understand the current topic under study. The current topic's content and files are under its dedicated directory. **Ignore** all other topic directories.

   2. **Read the topic-specific `README.md`**: (e.g., in `/mcp/`) to understand the specific mental models and technical constraints of the current subject.

   3. **Read `PROGRESS.md`**: To understand exactly where the last session ended and what the "Current Frontier" is.

   4. **Acknowledge State**: Start the session by briefly confirming the current learning milestone you've identified from the files.

Mostly, the agent concludes the session once it guages that 15-20 minute sprint-worthy concepts have been covered. Try it. If you want, I can share the full notes sans-my personal info through gist.

I'm a big fan of AI's ability to provide personalized tutoring.

So, lately, I have been using my Antigravity IDE (you can use any agentic harness) for personal learning. Things that bubble up during my daily work or just things I want to learn.

---

*Setup*

  learning/
  ├── AGENT.md
  ├── CURRENT.md
  ├── linux-admin/
  │   └── README.md
  ├── mcp/
  │   └── README.md
  ├── micro-gpt/
  │   └── README.md
  ├── postgresql/
  │   ├── PROGRESS.md
  │   └── README.md
  └── rag-v2/
      └── README.md

- CURRENT.md: current topic's directory path

- README.md: topic's learning goals, concepts I want to master

- PROGRESS.md: generated and edited by the agent on the fly

---

This is ACTIVE learning. As opposed to what Karpathy's LLM Wiki or any personal knowledge base does.

The learner undergoes 15-20 minutes learning sprints, tries things hands-on, asks to take notes, goes down the rabbit hole and agent tutor keeps the learning contextual and personal.

I have had an excellent experience with this setup so far.

Anyone doing this or anything similar?

Comments URL: https://news.ycombinator.com/item?id=47959169

Points: 4

# Comments: 4

While you research, this is the link to my article I had written a few weeks ago on my practical experience with the whole thing. I'm going to post it on HN today.

http://sisyphusconsulting.org/case-studies/2026/04/01/scalin...

I read about the alternatives. But has anyone tried anything that resembled Github Actions? Github Pages, I can take care of, but what about Actions? They're quite helpful to me in my day to day.

There should have been a two layer approach:

1. Run regex to screen for target word

2. If positive, run the context through a cheap AI model

One of the most fascinating parts to me was DNA transcription. The engineering is quite precise.

Found the video I was referring to: https://www.youtube.com/watch?v=7Hk9jct2ozY