Key takeaways
-
01
A meat proxy isn’t lazy. It’s latency.
-
02
Workslop costs its reader close to two hours.
-
03
We built Searle’s room and moved in.
-
04
“The AI said” is an alibi.
-
05
Ask: what did I change, and why?
[ ThruLine engineering · October 2026 · 12 min read ]
AI coding agents wrote most of ThruLine’s code, and ThruLine recorded them doing it. What its own record caught, what it got wrong about itself, and one fix we recorded end to end.
Ashwinsuriya Ravi · Founder
Read the research →
Key takeaways
-
01
1,537 of 1,941 commits name an AI co-author.
-
02
Its first finding on us was a false alarm.
-
03
The same rules caught our own mistakes.
-
04
A decision has to become a check.
-
05
We publish our failing rows.
[ ThruLine Research · October 2026 · 12 min read ]
CI checks the code an AI agent leaves behind. It doesn’t check what the agent did to get there, or whether what it told you is true. The evidence from 2025–2026.
Ashwinsuriya Ravi · Founder
Read the research →
Key takeaways
-
01
Green answers one question.
-
02
Models change the tests.
-
03
Passing isn’t correct.
-
04
The report isn’t evidence.
-
05
The worst failures never reach CI.