Telling AI 'don't' isn't a guardrail — build deterministic backstops
i TolD Ai NOT To dO SomETHing ANd iT diD it AnYwaY- people anthropomorphizing AI.
If that inevitability is not factored into your use of AI, you need to take a step back and find some tutorials… or perhaps didactic tales.
Doing AI-assisted development and think your “Don’t commit direct to main” instruction will save you? Go learn about proper git management before you’re responsible for a misunderstood or hallucinated instruction that brings your company to its knees.
The agentic workflow I’m working on now has deterministic limits built in beyond the prompt-based constraints for the orchestrator and throughout that *should* keep it well below those hard-coded caps. But "should" doesn't cut it.
I’m just working on the spec and prototypes, and I’ve already blocked Claude from touching main and am constantly adding in deterministic backstops and evaluations. This is not like handing something off to one of my trusted cofounders, a sprinkle of chaos is baked into these tools and I handle them accordingly.
Amazon can brute force their way through figuring out AI development and absorb the hit when things go awry. My and (likely) your company cannot.
Jessie Grenfell
Jessie has spent 20+ years architecting and delivering software for organizations where complexity is the norm — financial services, insurance, nonprofits, and regulated industries where getting it wrong isn't an option. She specializes in the hard problems: unique operational workflows, cross-system integrations, and the kind of compliance requirements (HIPAA, CCPA, SOC 2) that most developers treat as someone else's job.
At Codality, she leads every principal engagement from discovery through delivery — because she's seen what happens when that thread gets cut.