The amount of code written on local will tend to zero pretty rapidly from here. No one wants work to stop because the laptop got closed or an agent is waiting for a permission prompt. Almost all code will be written in remote sandboxes without any sort of supervision.
Agents change the cost structure of quantitative hedge funds. But this does not provide for a particularly good opening for startup managers.
andreasfragner.com/writing/a…
Being deep in the code is a necessary rather than a sufficient condition for designing well.
andreasfragner.com/writing/r…
Deciding not to solve a problem, or solving a different one, can also be a solution.
andreasfragner.com/writing/t…
Implementation — writing code — is a complement to specification — writing down what the code should do. Agents are commoditizing the former, pushing up demand for the latter.
andreasfragner.com/writing/a…
Hiring is an incomplete information game and to succeed you have to treat it as such.
andreasfragner.com/writing/h…
Andreas Fragner retweeted
Excited to share that Kappa has closed a $5M seed round led by Morningside! This investment will fuel our expansion across Africa and help us meet surging demand. Are you a Cameroonian business struggling with international payments? Get in touch!
kappapay.com/blog/seed-round…
High performers care deeply not just about doing the work but how it’s done. andreasfragner.com/writing/c…
Fewer parts make for better software and better products andreasfragner.com/writing/f…
Early on, you have to have the confidence to break with best practices. And that confidence comes from knowing what risks actually matter in your context. andreasfragner.com/writing/e…
When people ask for advice on personal investing, I’ve found they are either looking for confirmation that what they do is great (it’s usually not) or some sort of secret sauce for outperforming the market (which doesn’t exist).
andreasfragner.com/writing/i…
One way to think about software development is to view implementation as a complement to writing specs. If LLMs make the cost of the complement drop by, say, 90%, demand for the other may go up 10x — assuming there’s latent demand for the output. And for software there's plenty.
In the absence of a strong conceptual model for how LLMs work, following a systematic, empirical approach is the only way to get good results. Expect some cool tooling to be built around this. So prompt engineering moves from art to, well, engineering.
Prompt Engineering vs Blind Prompting. I show how prompt engineering can be a real systematic approach to extracting value from language models by following a realistic example, as opposed to "blind prompting" which is barely better than trial-and-error. mitchellh.com/writing/prompt…
The most rewarding things in life are hard but fun. Whether something hard is also fun is mostly a function of your attitude. Even easy things are painful if you take yourself too seriously. Viewing things as a puzzle or game might sound stupid, but it’s also a winning strategy.
“Just to future-proof things” is perhaps the most common argument I hear to defend questionable design decisions. Future-proofing only makes sense if you have some idea of what the future will look like and you understand what the tradeoffs are.
andreasfragner.com/writing/f…
I'm biased but this is a chance to work with an amazing team on some of the hardest and most rewarding problems in finance. You should apply.
Kappa is looking for superstar engineers with matching ambition to join our team. If you’re into building high-performance, high-availability systems with unique constraints, get in touch and apply below! We have both front- and backend roles open.
jobs.polymer.co/kappa
Jira is a fascinating case study in switching costs. The product has sucked for years but the pain is not strong enough to make me bite the bullet and switch to Linear or something else. Despite regularly cursing myself, still using it after all these years.



