1
Be specific about output shape
“Reply in one sentence” beats “be brief”. For property agents, say “Return a single value from this list: A, B, C”. The closer the prompt is to the coerced shape, the higher the success rate.
2
Focus and gate aggressively
A chat agent given a tight focus (scoped objects and pages) and a short list of allowed tools behaves more like a focused subject-matter expert than an unscoped general assistant. Scope is a soft hint (it biases the agent toward the right data rather than hard-blocking it) but fewer tools means fewer wrong turns.
3
Test on real records before relying on triggers
For property agents, run on five representative records manually before turning on automatic triggers (on create, on update). Verify the run trace, not just the written value.
4
Pick the right tier
Fast models (Claude Haiku 4.5, GPT-4o mini) are right for high-volume property auto-fill. Reasoning models (Claude Opus 4.7, o1-mini) are right for schema design and complex chat. Balanced models (GPT-4o, Claude Sonnet 4.6) are the safe default.
5
Watch the run trace
The property agent run panel shows what the AI produced and the value that was written. When something goes wrong, read the trace before tweaking the prompt blindly.