- One job, done well. Don't build "the everything agent". Build "the inbound SDR agent" and "the renewal-checkup agent" separately. They can hand off to each other.
- Define escalation crisply. "Escalate to human if: price discussion, contract terms, ‘talk to manager', or after 3 unresolved messages." Vague escalation = agent keeps trying when it shouldn't.
- Honesty over completeness. When the agent doesn't know, it should say so and offer to find out. Fabricated answers destroy trust faster than slow answers.
- Read transcripts weekly. The first month, read every conversation. After that, sample 10 per week. You catch drift, edge cases, and product gaps.
- Update the role doc, not the prompt. Treat the agent's role description as a living job-description. Edits there propagate to everything.
Best practices: scoping an agent that performs
Narrow scope, clear escalation, honest about limits — the three principles of agents that work.