Mark Müller

Co-founder and CTO at Working on monitoring and self-improvement for Software Factories. ML PhD from @the_sri_lab at @ETHZ

[Mark Müller @mnmueller]

This work was done with our https://t.co/wBU0bNbf0G. student Marco Paina @logic_star_ai, in collaboration with @the_sri_lab.

[Mark Müller @mnmueller]

New paper: Frontier Coding Agents Find Different Bugs Than Maintainers Fix

In proactive bug finding, agents report fewer than 20% of the bugs humans fixed. But they report a lot of other bugs, most of which are valid. This points towards an important distributional mismatch.

[Mark Müller @mnmueller]

If you want to learn more feel free to check out our website, blogpost, and paper!

Website: https://t.co/mERS7wU4GJ
Paper: https://t.co/fL2YJZUJ1d
Blogpost: https://t.co/bhKS9yzzR9

[Mark Müller @mnmueller]

@ptsankov Interesting observation. Could this be a sign of massive distillation of Western frontier models?

[Mark Müller @mnmueller]

10/10
This is the problem we work on at @LogicStarAI: monitoring and improving the real-world outcomes of coding agents run at scale.

[Mark Müller @mnmueller]

1/10
Went through Anthropic’s S-1. Two things stand out:

The unit economics look like normal software, at least for now
Over 90% of their own code is written by Claude, and the risk factors read like a list of what goes wrong at that scale 🧵

[Mark Müller @mnmueller]

9/10
Put together: generating code is cheap and getting cheaper, code per engineer is up 45x, agents can misreport “done”, and benchmarks don’t predict production.

The bottleneck has moved from writing code to knowing whether it is correct.

[Mark Müller @mnmueller]

@AnthropicAI Seems like we can get software margins after all:
Note 8 (p. F-38) cost of revenue: $2.4B on $16.3B of H1 2026 revenue, so ~85% gross margin. Just need to amortize that R&D over more tokens, $9.7B in H1, 60% of revenue