Collapsing three Opus calls into one isn't just a latency win — at ~$75/M output tokens, an always-on agent running 24 hourly cycles was burning through inference budget fast, so this is closer to a 66% cost cut on the LLM side per cycle. More interesting tradeoff is open-sourcing the prompt injection stack the same week you add trade routing via API — adversaries get the full defense playbook right as the attack surface expands from read-only news posting to actual execution capability. That said, public auditability on the security layers probably nets out positive given how many "AI agent" projects ship with zero input sanitization and wonder why they get drained.

Top comment by @Benthic