Hey everyone, I’ve run into something strange and wanted to see if anyone else has experienced this. We’ve built two AI agents using VAPI to place part orders by calling warehouse reps. - One agent handles website orders - The other handles eBay orders Both agents: - Use the exact same VAPI prompt - Pull identical part details, trims, and rep names from Airtable - Use the same phone number - Follow the same call structure and automation logic Yet somehow, the website agent is consistently performing better, while the eBay agent has more failed calls, even though the inputs are literally the same. We’ve already ruled out: - Trim formatting differences - Time-of-day patterns - Call volume differences - Assistant prompt mismatches - Session overlap or stale calls Everything looks clean but the eBay agent’s calls still fail more often (either reps hang up, get confused, or the call flow breaks). Has anyone seen something similar before? Could there be something subtle like session memory bleed, invisible differences in the way the call is initialized, or even rep behavior bias based on caller history? Would love to hear your thoughts or debugging tips 👇

Hey folks, I’ve been running into a bottleneck with Twilio, only a few of my outbound calls go through at a time, and the rest fail with a “Call concurrency limit exceeded” warning (Error 10004). I’m using Twilio for a legit voice automation setup where my assistant calls warehouse reps to confirm auto part availability. Just trying to get about 10 concurrent calls running smoothly. Anyone here successfully increased their outbound call concurrency with Twilio? Did you submit a request or go through a specific approval process? Would love to hear what worked for you! Thanks in advance!

Hey everyone I’ve been working on a fairly detailed voice assistant using Vapi, and I’ve noticed something strange: 👉 When the prompt gets too long or packed with logic, the AI starts ignoring parts of it — especially things like timing, waiting for replies, or skipping over instructions entirely. In my case, even though the prompt says “pause after asking how’s it going”, the assistant just jumps into the next line like it never read that at all. So here’s my question: Does prompt length affect performance? - Are we hitting some kind of internal memory/processing limit? - Should we break things into smaller behaviors or modular prompts? - Is there a best practice for structuring complex call flows? Would love to hear how others are handling large prompts in production. 🙏

Hey everyone 👋 I’ve been running into a pretty frustrating issue. I have my AI voice agent (Amelia from OEM Auto Space) set up with the dtmf_tool enabled in VAPI. The goal is for her to navigate phone menus (IVRs) and press the correct number to reach a human rep — always selecting the Sales option. But here’s the issue: 🔹 AI is not interacting with the IVR at all. 🔹 She keeps trying to speak to the rep while the automated menu is playing. 🔹 The result: no rep, no order — just stuck in IVR limbo 😔 The tool is definitely added under the Tools section, but I don’t see any sign that she’s recognizing or responding to the IVR prompts. Has anyone here gotten DTMF navigation working successfully in a real-world call? Any tips on config, tool instructions, or even a working example JSON would be super appreciated! Thanks in advance 🙏 — Ryan

Hey everyone — I’m running into a small but frustrating issue with my assistant on VAPI. In my prompt, I’ve clearly instructed the AI to: - Say “dash” instead of “minus” for hyphens (like in LJ6T-18B955-FD) - Speak IDs slowly, one character at a time - Say “zero” instead of “O” - Only apply this behavior when reading part IDs inside the trim Despite all that, Amelia (the AI) still occasionally says “minus” instead of “dash,” especially when the trim is read automatically (not after being asked). It’s inconsistent — sometimes she gets it right when I ask for the ID, but not when she reads the full trim. Has anyone found a reliable fix for this behavior? Would love any advice or prompt strategies that worked for you. Thanks 🙏