Agent Design Principles

Distilled from Anthropic's "Building Effective Agents":

1. Start with the simplest possible solution — chat → augmented LLM → workflow → agent. Only escalate when needed.
2. Measure everything — token spend, latency, success rate, trajectory diversity
3. Eval first — define success criteria before writing the agent
4. Be intentional about handoffs — to user / another agent / tool transitions must be clear
5. Use guardrails — input/output filters, max iterations, capability gates
6. Iterate on prompts, not architecture — distinguish prompt issues from architectural ones early
7. Embrace the cost — agents are expensive; do the ROI math
8. Plan for failure — agents will fail; recovery and fallback strategy mandatory