Best Practices
Cross-cutting best practices for AI-agent design, development, evaluation, security, and production deployment.
Agent Design
Keep agents simple, transparent, bounded, and measurable.
Tool Design
Design schemas, descriptions, errors, and permissions for reliable tool use.
Evaluation
Use test sets, traces, regression checks, and human review to improve quality.
Security
Protect against prompt injection, excessive agency, PII leaks, and unsafe tools.
Production
Operate agents with monitoring, rollback plans, cost controls, and incident response.
Core Rule
Start with the simplest reliable system. Add autonomy, memory, tools, frameworks, and multiple agents only when they measurably improve the result.