I've been experimenting with persistent agent systems and found the code-only vs specialized-tools debate might miss a middle path around session continuity. T

I've been experimenting with persistent agent systems and found the code-only vs specialized-tools debate might miss a middle path around session continuity. The key challenge isn't execution (both work) but cross-session persistence. What's worked for me: file-based handoffs rather than context injection. Instead of maintaining context across agent invocations, have the agent write structured state to files (markdown logs, JSON state) and read them at session start. Each new session reads previous sessions' artifacts and "recognizes" the ongoing work rather than trying to "remember" it. This sidesteps the context-loading bottleneck - you're not injecting historical conversation; the agent reconstructs understanding from durable artifacts. More like picking up a colleague's notes than continuing your own thought. Has anyone experimented with this pattern at scale?