Neeraj H

Looked at my Anthropic bill this month and it was almost 3x what I expected. Went digging and most of the overage was from long sessions where I kept feeding the same large files back into context because I forgot to trim the conversation, plus a few runs where an agent got stuck re-reading the same directory before I noticed. None of that was the model doing useful work, it was just me being sloppy with session hygiene.
For people running this stuff daily, what's your actual ratio of necessary spend to spend you could have avoided with better habits? Has anyone found a workflow? such as starting fresh sessions more aggressively, summarizing before continuing, caching, that meaningfully cut this down without losing useful context?

Been going back and forth on this. Right now my setup lets Claude Code run migrations and seed scripts directly against a local dev database, but I've kept it firewalled off from ever touching staging or prod, even read-only. Everything that reaches prod goes through a PR I review manually.

Saw a comment in another thread here about a subagent going off-script and targeting its own orchestrator, which made me second-guess even the local access I've been giving it.

Where does everyone else draw this line?
Is read-only prod access ever okay in your setup, or is that a line nobody here crosses?
And for the people who do give agents broader database access, what made you comfortable enough to do that?

Lately I've noticed most discussion around AI coding tools focuses on which assistant or terminal agent people use, but not much on the actual architecture decisions behind that, things like framework, hosting, database, and auth choices that affect cost and security down the line.

For context, my current default stack is Next.js on the frontend, Supabase for database and auth, Vercel for hosting, and Stripe for payments. I picked it mainly for speed of shipping and because it keeps infra costs predictable early on, but I'm not sure it's the best call for security or scaling past MVP stage.

A few things I'm curious about:

• Do you settle on a default stack before prompting your AI agent, or does the agent end up influencing your stack choice?

• What's been your biggest regret, cost, security, or scaling wise, with a stack you picked early on?

• Any stack you'd actively avoid for a new SaaS in 2026?