I'm good at claudemaxing.
I’ll be honest: I’m more of a vibe coder right now than a proper engineer.
For the last few weeks I’ve been leaning heavily on AI agents (mainly Claude-style “do X for me” workflows) to build and ship. On paper it sounds like a cheat code: describe the feature, let the agent write the code, glue it together in VS Code, ship.
In practice, a few things hit me hard:
- Agents are great at getting you from 0 → 60%, but terrible at owning the last 40%. That’s where all the edge cases, race conditions, and weird state bugs live.
- When something breaks, you realize you don’t fully understand the codebase the agent produced. Debugging turns into “prompt archaeology” instead of actually reasoning about the system.
- For anything touching money (billing, credits, limits, retries, etc.), “agent did it” is not a real safety net. You need to know exactly what’s happening and why.
My takeaway so far:
- Agents are amazing for scaffolding: boilerplate, wiring up basic CRUD, even drafting docs and onboarding flows.
- But for core product logic and anything in the payment / credit / auth paths, you have to treat them like untrusted juniors. Review everything, simplify aggressively, and rewrite where needed.
- If you don’t have at least a basic mental model of your own stack, you’re not a SaaS founder, you’re just running prompts on top of a black box.
I’m not “anti-agent” – I still use them every day. I’m just realizing they’re not something you can fully outsource your brain to yet, especially if you’re handling real users, real money, and real risk.
Curious if anyone here is actually running a production SaaS where agents own meaningful chunks of the codebase long term, or if most of us are quietly rewriting their work once things get serious.