Humans learn proprietary wonkiness. Gen AI burns tokens on it

Humans are good at learning and accommodating proprietary wonkiness. Gen AI is not. That makes Salesforce inherently problematic as a substrate for agentic development.

A seasoned Salesforce professional knows where the mines are buried. Objects that follow different SOQL rules. Standard fields that won't map to custom fields on Lead conversion. I'd get carpel tunnel trying to list them all.

Most of those gotchas are documented, either in Salesforce's own docs or in the laments of frustrated admins on forums. So you have a choice: pre-load what you know will be relevant into skill files, or leave let your LLM development harness resolve gotchas as its logic gets slapped by them.

The problem with the first option is context collapse. GenAI accuracy drops as the context window fills with your existing code, your new request and a tome of rules to follow. The harness trial-and-error method can be messy with years of now deprecated solutions floating about, not to mention the people compelled to answer questions with irrelevant solutions.

Either path, you're looking at token burning, hallucinations and risk.

With my work flipping from full stack apps to Salesforce by project these days, I feel the difference. I find myself stopping the harness the moment I recognize it's wrestling with a Salesforce gotcha and point it in the right direction. The harness advances are impressive. They'll brute force their way to a solution if you let them. But you'll blow through your usage budget while the tool tries every loosely related fix it scraped off Reddit.

Full stack apps hit wonky issues too. Far less frequently though, and the solution is often quickly identified. That's not Salesforce's burden alone. Legacy baggage leaves a lot of the OG code-customizable platforms at the same disadvantage.

The interesting twist? Structured, consistently abstracted configuration is exactly what GenAI tooling handles well, the same disciplined patterns engineers use to keep code sane. The very thing Hype Lords have declared GenAI would make obsolete.

I regret to inform you that the priest class skills of software developers, so often dismissed as overengineering by the Dunning-Kruger cult, are more valuable than ever.