How to keep AI agents working from current project information
A thread on the stale-context problem: agents working from last week's plan confidently do last week's job. The pattern that works is treating project docs, tickets, or the repo as the source of truth that the agent re-reads at the start of every run, plus a one-page current-state file updated after each session. Fresh context beats a bigger memory, because an agent's memory is exactly where old information hides.