jghiglia

I've been using Claude Code heavily for the last few weeks building out a multi-agent system, and the token economics caught me off guard — I hit 75% of my Pro weekly budget faster than expected. I don't code myself, so Claude Code handles all the actual implementation work.

What I've learned about cost management: the real decision isn't "should I code this myself or use Claude Code" — it's "should I spawn Claude Code or handle this through a different approach entirely?" For complex builds where I need something architected from scratch, Claude Code is worth it. For smaller tasks or iteration on existing code, I've started using Sonnet in the web interface with the file as context instead. The visual guide here is helpful because it shows you what Claude Code is actually doing under the hood — understanding that workflow helps you predict whether a task will be a quick fix or a deep exploration that burns through your budget.

SQLite works well for single-machine single-user. The constraint we kept hitting is portability — switching between Claude Desktop, Claude Code, and Claude.ai means three different surfaces that can't share the same local SQLite. Also breaks down for multi-agent setups where Agent A needs Agent B's memories. We went hosted for exactly this reason. Local is a great starting point but the ceiling is low once you're running more than one agent or more than one device. What's your migration path if someone needs to move off local storage?

LoCoMo is a good benchmark but it tests single-agent recall. The harder problem for production multi-agent systems is cross-agent correction — when Agent A learns that a fact stored by Agent B is wrong. Does Cortex have a propagation mechanism for that or is correction handling left to the application layer?

The self-editing angle is interesting but the part that breaks in production is what happens when the search agent's memory of previous searches is wrong or stale. We ran into this building multi-agent systems — the agent confidently re-runs searches it already ran because it has no reliable memory of what it tried. The self-editing loop helps but only if the memory layer underneath it is trustworthy.