@Matan_G_ Much of it is archeology ("why did we add this very specific validator?"), which then lets you build falsifiable eval cases for that failure mode (ideally replaying prod traces), so you can run it on the new arch and see if it needs that lesson ported
Nothing like a fresh start. Models move so fast that domain-specific harnesses accumulate obsolete code within weeks. How do you enjoy all the benefits of rewriting from scratch without losing hard-earned operational knowledge embedded in code? >>
There’s an incredible amount of technical depth behind building AI agents that can actually work for executives.
We’re building an AI agent that needs to be reliable, accurate, smart, and fast, while also adapting extremely quickly to the individual working style of each executive.
@Dorbernsohn and the team are doing brilliant work on this!
Read more here:
https://t.co/xoB81ifvLJ
@adambn Indeed. Using isolated subagents is powerful because it still allows the main agent to orchestrate communications on untrusted channels with a much smaller boundary to protect
You may know subagents as a way to maintain an efficient, modular context, fan out concurrent work, etc. But they could also be a powerful tool in keeping minimal trust boundaries when building agents that communicate externally.
https://t.co/SVugJmDn1F
@theDushi1 אתה מאתחל סשן של קלוד קוד עם bypass all permissions, שעולה אוטומטית, עם סקילז שאומרים לו לקרוא תוכן שעלול להיות זדוני בלי שום הגנות נגד prompt injection, וספציפית הוא אפילו יכול לערוך את ההגדרות של עצמו
@vincent_koc Not only that but @AnthropicAI have been ignoring my open issue about CC pushing a line for each worktree into the sandbox profile making it unusable entirely: https://t.co/klz87uFsNg
@bradwestfall@Tomas_Zubiri Definitely agree! I wanted to highlight the fact that not only is it never satisfied, it actively degrades a valid solution when you tell it to improve it