Post

The underrated aspect of agent harness engineering: after an agent fails, what becomes permanent? A new rule? a test? a tool permission? a memory? a source link? nothing? Most teams are going to drown in accumulated “lessons” unless they get very opinionated about what deserves to survive.

12/5/26, 6:01 pm0 reposts0 likes

josh

@joshpham.com

12 May

For people building agents: when something goes wrong, do you record it as: - eval case - prompt/rule change - memory - tool/hook - human process - “ignore, one-off”