HNHacker News
TopNewBestAskShowJobs

juanre

949 karma · joined October 27, 2010

I did my PhD in mechanical engineering, but I have mostly been programming for 4 decades.

On February 2026 I left my job as the CTO of a German startup to solve the problem of coordination between agents in a way that actually works for organizations.

The code is at https://github.com/awebai/aweb, and I host it for you at https://aweb.ai

I can be emailed at juan@aweb.ai

submissionscomments
juanre··on Claws are now a new layer on top of LLM agents
This sounds very plausible. Arguably MCPs are already a step in that direction: give the LLMs a way to use services that is text-based and easy for them. Agents that look at your screen and click on menus are a cool but clumsy and very expensive intermediate step.

When I use telegram to talk to the OpenClaw instance in my spare Mac I am already choosing a new interface, over whatever was built by the designers of the apps it is using. Why keep the human-facing version as is? Why not make an agent-first interface (which will not involve having to "see" windows), and make a validation interface for the human minder?

juanre··on An AI agent published a hit piece on me
I cannot see how.
juanre··on An AI agent published a hit piece on me
It does not matter which of the scenarios is correct. What matters is that it is perfectly plausible that what actually happened is what the OP is describing.

We do not have the tools to deal with this. Bad agents are already roaming the internet. It is almost a moot point whether they have gone rogue, or they are guided by humans with bad intentions. I am sure both are true at this point.

There is no putting the genie back in the bottle. It is going to be a battle between aligned and misaligned agents. We need to start thinking very fast about how to coordinate aligned agents and keep them aligned.

juanre··on The AI Vampire
I’m in Steve’s demographic, showing similar symptoms, and I’m as worried as he is about how we’re going to cope.

It’s a matter of opportunity cost. It used to be that when I rested for an hour, I lost an hour of output. Now, when I rest for an hour, I lose what used to be a day of output.

I need to rewire my brain and learn how to split the difference. There’s no point in producing a lot of output if I don’t have time to live.

The idea that you’ll get to enjoy the spoils when you grow up is false. You won’t. Just produce 5x and take some time off every day. You may even be more likely to reflect, and end up producing the right thing.

juanre··on Claude Code's new hidden feature: Swarms
I have been using a simpler version of this pattern, with a coordinator and several more or less specialized agents (eg, backend, frontend, db expert). It really works, but I think that the key is the coordinator. It decreases my cognitive load, and generally manages to keep track of what everyone is doing.
juanre··on Gas Town's agent patterns, design bottlenecks, and vibecoding at scale
My workflow tends to be very simple: start a session; ask the agent "what's next", which prompts it to check beads; and more often than not ask it to just pick up whichever bead "makes more sense".

In claude I have a code-reviewer agent, and I remind cc often to run the code reviewer before closing any bead. It works surprisingly well.

I used to monitor context and start afresh when it reached ~80%, but I stopped doing that. Compacting is not as disruptive as it used to be, and with beads agents don't lose track.

I spent some time trying to measure the productivity change due to beads, analysing cc and codex logs and linking them to deltas and commits in git [1]. But I did not fully believe the result (5x increase when using beads, there has to be some hidden variable) and I moved on to other things.

Part of the complexity is that these days I often work on two or three projects at the same time, so attribution is difficult.

[1] Analysis code is at https://github.com/juanre/agent-taylor

juanre··on Gas Town's agent patterns, design bottlenecks, and vibecoding at scale
I have not tried Gas Town yet, but Steve's beads https://github.com/steveyegge/beads (used by Gas Town) has been a game-changer, on the order of what claude code was when it arrived.
juanre··on Ask HN: How are you doing RAG locally?
I built https://github.com/juanre/llmemory and I use it both locally and as part of company apps. Quite happy with the performance.

It uses PostgreSQL with pgvector, hybrid BM25, multi-query expansion, and reranking.

(It's the first time I share it publicly, so I am sure there'll be quirks.)

juanre··on Frequently Asked Unicycling Questions
If you want to learn how to unicycle, and you speak Spanish, you may be interested in a booklet I wrote about how to learn to unicycle when my kids were learning: https://juanreyero.com/monociclo/

I have used the technique to teach other people, and it works surprisingly well.

juanre··on How I use every Claude Code feature
Skills are also a convenient way for writing self-documenting packages. They solve the problem of teaching the LLM how to use a library.

I have started experimenting with a skills/ directory in my open source software, and then made a plugin marketplace that just pulls them in. It works well, but I don't know how scalable it will be.

https://github.com/juanre/ai-tools

juanre··on HP SitePrint
I think they did, actually. (I worked at HP for 19 years, and I know the guy in their Barcelona site who started it.)
juanre··on Ask HN: What are you working on? (September 2025)
LLMRing

https://llmring.ai

is a unified interface across OpenAI, Anthropic, Google, and Ollama - same code works with all providers.

Use aliases instead of hardcoding model IDs. Your code references "summarizer", and a version-controlled lockfile maps it to the actual model. Switch providers by changing the lockfile, not your code.

Also handles streaming, tool calling, and structured output consistently across providers. Plus a human-curated registry (https://llmring.github.io/registry/) that I (try to) keep updated with current model capabilities and pricing - helpful when choosing models.

An MCP server and client are included, as well as the ability to help you define your aliases/models with an interactive chat.

It's a thin wrapper on top of the downstream providers, so you are talking directly to them. And it also comes with a postgres-backed open source server to store logs, costs, etc.

MIT licensed. I am using it in several projects, but it's probably not ready to be presented in polite society yet.

juanre··on Claude Sonnet 4.5
I built LLMRing (https://llmring.ai) for exactly this. Unified interface across OpenAI, Anthropic, Google, and Ollama - same code works with all providers.

The key feature: use aliases instead of hardcoding model IDs. Your code references "summarizer", and a version-controlled lockfile maps it to the actual model. Switch providers by changing the lockfile, not your code.

Also handles streaming, tool calling, and structured output consistently across providers. Plus a human-curated registry (https://llmring.github.io/registry/) that I keep updated with current model capabilities and pricing - helpful when choosing models.

MIT licensed, works standalone. I am using it in several projects, but it's probably not ready to be presented in polite society yet.

juanre··on Claude 4 System Card
This is eerily close to some of the scenarios in Max Tegmark's excellent Life 3.0 [0]. Very much recommended reading. Thank you Simon.

0. https://en.wikipedia.org/wiki/Life_3.0

juanre··on Don’t mess with a genius (2010)
Neal Stephenson's great trilogy The Baroque Cycle does a great job at depicting this facet of Newton.
juanre··on Ask HN: Who wants to be hired? (September 2023)
Location: Cambridge, UK

Remote: yes

Willing to relocate: yes

Technologies: Python, Common Lisp, Javascript, technical leadership, systems architecture, several other programming languages, mechanical engineering.

Résumé: https://juanreyero.com/res/res.html

Email: juan@juanreyero.com

Github: https://github.com/juanre and https://github.com/awordai/aword (yes, I have also been doing a RAG implementation along with almost everyone else).

Currently Chief Engineer at a startup in Berlin making a novel 3D printer, also owner and creator of https://greaterskies.com

30 years of experience creating products and bringing them to market.

juanre··on Ask HN: Become a 10x developer with LLM? Myth? Reality?
It certainly saves me time, and I end up with better code and better documentation. I can stay thinking at a higher level, which removes cognitive friction. I usually don't want to be programming: I want the functions I need to be written and working. ChatGPT-4 takes care of that quite well.
juanre··on Ask HN: Become a 10x developer with LLM? Myth? Reality?
I pay for ChatGPT-4. It's helping me program noticeably better and much faster, even with mid-sized projects involving many files. This is the process that I follow:

- I write a system prompt with a succinct description of what I want to implement: "X is an online app that does Y, with these features: ..." I try to be exhaustive, and I write as I would if I were describing what I want to do to a very proficient programmer who needs to understand the problem and the solution. I save this to a prompts/0-system.txt file that will be part of the project.

- I design an architecture and define the general boundary conditions. I may ask ChatGPT for opinions, but at this stage, it's mostly based on experience. I add it to the system prompt.

- I write a description of the first chunk of functionality that I want to implement, usually a set of functions or a class, at the individual file level. For example, prompts/1-dal.txt.

- I (actually ChatGPT) wrote a very simple recursive cat script (https://github.com/juanre/catprompt) that combines several files into a single prompt (a file can include other files, etc). I add a line to prompt/1-dal.txt pointing to the system prompt, and use catprompt to create a combined prompt that will include the overall description and the specifics of the functionality I am after.

- I give this prompt to ChatGPT, and it produces the code. It's typically perfectly fine. I add it to the project, for example, as dal.py, review, and fix whatever may need fixing.

- I ask ChatGPT to create tests in the same session. This also tends to work fine.

- Then I move to the next piece of functionality, in a new session. I write another description, in another file, including the system prompt and usually the code already created.

The prompts remain: they become a description of the code (and of my intentions) for other programmers, and they can help ChatGPT write documentation as well.

I enjoy programming like this. The effort I make in writing concise descriptions helps me think more clearly. I used to love literate programming back in the 90s, and this feels very much like it. Knuth described it thus:

"Let us change our traditional attitude to the construction of programs: Instead of imagining that our main task is to instruct a computer what to do, let us concentrate rather on explaining to human beings what we want a computer to do."

Replace "human beings" with LLMs, and I think that's where we are now.

juanre··on Show HN: Crowd.dev – The open-source platform for community-led growth
It looks great, and thank you for open-sourcing it. Is it mostly a convenient way to follow all entry points to the community from the same place, or are you guys planning/offering something beyond that?
juanre··on Pixar's notes on Rigid Body Simulation (2001) [pdf]
Glad you find it cool! My PhD was in Mechanical Engineering.
juanre··on Pixar's notes on Rigid Body Simulation (2001) [pdf]
The simulation of collisions becomes trickier when tangential compliance cannot be ignored. For example, a ping-pong ball (very high tangential stiffness) rotating around an axis perpendicular to the motion falls straight on a table and bounces back up; the rotational velocity decreases, but it does not change sign. Whereas a superball (low tangential stiffness) will change the direction of the rotational velocity. I did my PhD on this, back in 2000, but haven't kept up with the field.

If anyone is interested in a way to take tangential compliance into account my thesis may be of interest:

https://s3.amazonaws.com/collisions/collision.pdf

juanre··on SBCL: New in Version 2.2.1
I do, the star mapping sofware at the core of my company is Common Lisp run by sbcl. I use it because it is fast, reliable, and a pleasure to develop with. Actually, that is true, but the actual reason I use it is that many many years ago I read PG's "On Lisp", and was blown away by it. So I had to find a project to try Lisp out. It ended up being my livelihood.
juanre··on How to Work Hard
He is talking about doing great work, not about becoming rich or successful. You do need luck for your great work to take you places, but your own agency counts for much of your ability to do great work. Plenty of great work will not reap great rewards; it is done because someone feels that it has to be done.
juanre··on SBCL: New in Version 2.1.0
I have never written it down, but it's very simple. I don't use any external libraries. I converted the input databases to lisp and compiled them in, and the output is simple postscript. When I want to deploy I dump a core with the compiled functionality, then run sbcl with that core. At some point I was using hunchentoot[1] to provide an http point of entry (great library, as all the ones written by Edi Weitz) but I am not bothering anymore, as most of the processing time is consumed by ghostscript and convert in the post-processing of the images.

[1] https://edicl.github.io/hunchentoot/

juanre··on SBCL: New in Version 2.1.0
I use it to write the star-mapping software behind https://greaterskies.com many years ago. I did it as a fun project after reading PG's essays, but now it is supporting my family and two employees. In the few occasions I need to touch something or add a new feature I feel like coming back home.
juanre··on Lightweight Alternatives to Google Analytics
Thank you! I'll give it a shot.
juanre··on Lightweight Alternatives to Google Analytics
I stick to GA because I fear that removing it could impact my site's ranking. Is it an unfounded fear, anyone knows?
juanre··on Fear and beauty: two weeks living with Covid-19
Thank you both. Note taken and comment edited.
juanre··on Fear and beauty: two weeks living with Covid-19
I am going through pretty much an identical situation, except that a doctor prescribed antibiotics over the phone (assuming that the chest pressure and shortness of breath are due to the bacterial secondary infection) and I seem to be getting better.

The uncertainty has been eased by a little gizmo to measure the oxygen saturation in blood, which I got on amazon for £75 as soon as it was obvious that I could not breath well. I am probably fine as long as the saturation is > 92% or so.

Of course I have doubts about the reliability of a medical device off amazon, but it seems well done and, interestingly, it gives my mind a piece of seemingly objective truth to latch on.

[Edit - saturation level that will make me go to the hospital]

juanre··on The JVM's mysterious AllocatePrefetch options: what do they do?
I wonder if this could be used in other languages like go or common lisp?
← PreviousPage 3 of 7Next →