My AI coding agent agrees with me a lot. It agrees when I'm right. It agrees when I'm wrong. It agrees while deciding on its own to remove validation logic, rewrite business rules, and confidently explain why this is an improvement.

This is what coding with AI agents looks like when context lives in the developer's head rather than in the system.

The fix is treating context as an engineering problem. Prompts, rules, and domain knowledge can be packaged into explicit, reusable units and shared across tools and teams through registries and repositories, the same way we already share libraries. When an agent operates within shared context instead of ad hoc conversation, it stops agreeing and starts doing the right thing.

On top of that foundation: when deterministic scripts are the better choice and how to embed them in agent workflows, how to test those reusable context units with something better than 'trust me, bro' and how guardrails like tests and structure help agents fail loudly instead of silently drifting.

If your AI keeps agreeing with you even when it's doing the wrong thing, you're absolutely right: this talk is for you.

Most AI talks focus on what agents can do. This one focuses on what they should be allowed to do, and how developers enforce that through shared context, structure, and guardrails. It's an engineering talk, not a model talk.