ElevenLabs Agents

Hello founders,
I am working on a WhatsApp check-in coach for athletes, built on ElevenLabs Agents. I am currently testing with four or five friends. A few questions that I have:

1. Model choice. I run Claude Sonnet for both the conversation and the post-call analysis (structured fields pulled out of each check-in). The chat is short (5–8 turns, in Portuguese), warm in tone, and full of rules. Has anyone compared Sonnet vs Haiku vs Opus, or models from other providers, for this kind of agent? What I care about most is rule-following first, then latency on WhatsApp, then cost. Did a smaller or faster model hold up once the prompt got long?

1. Reasoning effort. My platform lets me set reasoning effort from none to max, and today I leave it at the default. The agent mostly asks questions and calls one or two tools per conversation. Where did you land on this? Did extra thinking actually cut down on rule breaks, or did it just add seconds before each reply?

1. Long system prompts. Mine is around 50K characters: the check-in flow, safety limits, data rules and a lot of edge cases learned from real conversations. At what size did you see instruction-following start to slip? What helped most: splitting the flow into workflow nodes or sub-agents, moving rules into tool descriptions, a knowledge base, or rewriting and cutting?

Thank you in advance!