aws-samples Cost Optimization for AI Agents
Model Selection
- Claude Haiku for simple routing/classification (~10x cheaper than Opus)
- Claude Sonnet for most agent tasks (good balance)
- Claude Opus only for complex reasoning/code generation
- Amazon Nova Lite/Micro for high-volume, simple tasks
- Consider prompt caching for repeated system prompts (up to 90% savings)
AgentCore Pricing
- Runtime: per-second compute (scale-to-zero = no idle cost)
- Memory: per-event storage + per-retrieval calls
- Browser: per-session charges
- Network: standard EC2 data transfer rates
Optimization Strategies
- Prompt engineering: shorter prompts = fewer tokens = lower cost
- Lazy skill loading: only load full SKILL.md when needed (not in system prompt)
- Memory compaction: CompactionHook to summarize long conversations
- Session pooling: reuse agent instances across invocations
- Tiered model routing: use cheap model for triage, expensive for complex tasks
- Caching: cache tool results (e.g., API calls) to avoid redundant work
Memory Cost Control
- STM events: minimize event size (don't store large payloads)
- LTM strategies: 4 strategies × extraction cost per event
- Retrieval: semantic search has per-query cost
- Tip: use explicit save_memory for important facts only, don't over-index
Monitoring Costs
- CloudWatch metrics for invocation count, duration, token usage
- Set billing alarms for unexpected spikes
- Track per-user usage for chargeback models