Will explore this for https://toolkami.com/, which allows plug and play advanced “code mode” for AI agents.

Seems like it can be a standard the project can adopt.

I use headless codex exec a lot, but struggles with its built-in telemetry support, which is insufficient for debugging and optimization.

Thus I made codex-plus (https://github.com/aperoc/codex-plus) for myself which provides a CLI entry point that mirrors the codex exec interface but is implemented on top of the TypeScript SDK (@openai/codex-sdk).

It exports the full session log to a remote OpenTelemetry collector after each run which can then be debugged and optimized through codex-plus-log-viewer.

codex-plus provides a CLI entry point that mirrors the codex exec interface but is implemented on top of the TypeScript SDK (@openai/codex-sdk).

It exports the full session log to a remote OpenTelemetry collector after each run which can then be debugged and optimized through codex-plus-log-viewer.

Have a look at https://github.com/aperoc/codex-plus!

Comments URL: https://news.ycombinator.com/item?id=46644744

Points: 1

# Comments: 0

Plenty of engineers are struggling with their identity in this new age. Anxiety shows up even among the best of us, and anger isn’t far behind. - Andrej Karpathy (https://x.com/karpathy/status/2004607146781278521) - Rob Pike (https://itsfoss.com/news/rob-pike-furious/)

I know both emotions. I pride myself as a software craftsman, yet I am also the co-founder and CTO of an AI startup that keeps getting run over by paradigm shifts and better-funded startups (story for another day). I now know better the paths to stay out of; what matters more is knowing how to pick myself up and go again in this new age.

To get out of that spiral, I had to change what I think my value is.

First, our value is not in writing more code.

This isn’t new, and good engineers have been saying it for a long time: The best code is no code at all (https://blog.codinghorror.com/the-best-code-is-no-code-at-all/).

Code is a liability, not an asset. Every line is a future maintenance burden. Every new feature expands the surface area for bugs.

In today’s environment, more code also means more AI context, which leads to degraded performance.

Our value lies elsewhere, and John Carmack said it clearly: “Coding” was never the source of value, and people shouldn’t get overly attached to it. Problem solving is the core skill. The discipline and precision demanded by traditional programming will remain valuable transferable attributes, but they won’t be a barrier to entry. - https://x.com/ID_AA_Carmack/status/1762110222321975442

So do not be afraid to throw code away. Your discipline and problem-solving skills stay with you. I have thrown away more code in the last 2 years than I ever imagined.

More importantly, the goal is to create value for others, and very little of that is pure intelligence.

If intelligence were everything, the world would be run by people with 250 IQ. It is not.

AI has a narrow kind of intelligence. It works well only when the problem looks like this, as Demis Hassabis notes in his Nobel Lecture: 1. Massive combinatorial search space 2. Clear objective function (metric) to optimise against 3. Either lots of data and/or an accurate and efficient simulator

But real work is messy and unique: - The person you are helping is not you, and does not share your strengths, weaknesses, or resources - The times and context in the training data are different from yours - And so on When those three conditions aren’t true, AI looks smart but still fails at basic, real-world decision-making. Anthropic’s vending-machine experiment shows how much still depends on experience, intuition, and real-world constraints.

Linus Torvalds has the same sentiment on intelligence: And don’t EVER make the mistake that you can design something better than what you get from ruthless massively parallel trial-and-error with a feedback cycle. That’s giving your intelligence much too much credit.

Ruthless feedback beats raw intelligence. That is the core of high agency.

That feedback loop is what high agency looks like in practice. The fastest path to user value is a short feedback cycle, from information to action, for you, your team, and your AI.

Observe the people around you, figure out what they hate doing, learn to do it, and take it off their plate through software. In the long run, the highest net value creator wins.

If it is a problem well suited for AI, build data pipelines or design a simulator for it. Let them take care of it and move on to higher value problems.

AI gives us, especially software engineers, the ability to make our ideas a reality again and again to generate better ones over time.

Be prepared to throw away a lot of code, because the loop—observe, decide, ship, learn—is the value. We’re just getting started.

Please share your thoughts!

Comments URL: https://news.ycombinator.com/item?id=46537469

Points: 12

# Comments: 1

[0]: https://blog.toolkami.com/mcp-server-in-a-file/

SWE-grep was able to hit ~700tokens/s and Cursor ~300token/s, hard to compare the precision/recall and cost effectiveness though, considering SWE-grep also adopted a "hack" of running it on Cerebras.

I'm trying to kickstart a RL-based code search project called "op-grep" here[1], still pretty early, but looking for collaborators!

[0]: https://cognition.ai/blog/swe-grep [1]: https://github.com/aperoc/op-grep

I mean it's not much, but the concept just resonates with me and I want to share it. Sad I can't share even simple opinion nowadays ...

Comments URL: https://news.ycombinator.com/item?id=45728059

Points: 6

# Comments: 15

[0]: https://blog.toolkami.com/alphaevolve-toolkami-style/

[0]: https://github.com/tobi/try/blob/main/try.rb