When AI deployments stall, it's almost never because leadership lacks vision. It's because someone over-engineered the Flow — or forgot that agents don't actually "learn" from feedback unless you train them to.
Salesforce FDE Conner Hobbs recently dropped a piece that's required reading if you're deploying Agentforce in earnest. He's the kind of voice you only get from someone living in customer orgs day in and day out, fixing the exact failures that don't make it into the marketing deck.
What FDEs see that you don't
- Hallucinations
- Latency issues
- "Smart" agents that don't actually do anything useful
They've been there. They've fixed it. The lessons are repeatable.
Inside Conner's breakdown
- How over-engineered Apex and broken token flows nearly wrecked a top jewelry brand's agent — and the cleanup pattern
- Why structured prompts usually fail — and the simpler approach that actually works
- The truth about agent "memory" — hint: you have to build it
- When to use LLMs vs Flows vs Apex — and, more importantly, when not to
These aren't best practices. They're battle-tested truths from the people fixing your pipeline, prompts, and latency before your execs panic.
If you're a RevOps or IT leader deploying AI in Salesforce, this is gold.
Read Conner's full post. And if you want to build at this level, the FDE team is hiring.