Julian Neagu

The small context packet idea is underrated. I’ve seen agents get worse just because we keep feeding them everything instead of what they actually need.

The three-layer split makes a lot of sense. I’ve hit this with agents too — session notes pile up fast, but the decisions that matter are usually the “why”.

The shared table example is exactly where AI feels useful to me. The code can look fine while the real problem is hiding in another service.

The session-to-user binding bit stood out to me. It’s easy to assume the sandbox solves everything, but identity can still be the weak link.

The batch vs interactive split is probably the most useful takeaway here. For background jobs, I’ll happily trade a bit more time for a much lower bill.

The 91%+ input cost really stood out to me. I've seen the same thing with agent tools where the context quietly costs more than the actual answer.

The part about RAG being a search problem, not just a vector DB problem, hits home. Exact IDs and dates can matter more than semantic similarity. Learned this the hard way.

The HTTP 200 point is the bit that gets missed. I've seen AI tools look healthy on paper while the actual output was quietly getting worse.

The part about permissions being enforced outside the prompt is key. If the agent can't call the tool, there's nothing to argue about.

I like the tradeoff here. For agent mail, owning the inbox feels way better than adding another hosted service to trust. Cloudflare Workers is a pretty natural fit for this too.