Choose your stack: platform, framework or realtime model

Intermediate⏱ 13 min📝 Quiz
A decision guide for picking the right build path — and how to avoid being locked into a tool that changes next quarter.

Outcome of this lesson

You pick a build path for your first agent using clear criteria, not hype.

Three build paths

PathExamples (names change — verify current options)Best when
Voice-agent platform (hosted, low-code)Vapi, Retell AI, Bland, ElevenLabs Agents, SynthflowYou want a working phone agent in a day; teams, agencies, pilots.
Open framework + your cloudLiveKit Agents, Pipecat, Twilio Media Streams + your codeYou need custom logic, data control, lower cost at scale, or special compliance.
Speech-to-speech realtime APIRealtime audio models from major AI labsYou prioritise natural, low-latency conversation and accept less step-by-step control.

Decision criteria (score each 1–5)

  1. Speed to first call — hours vs weeks.
  2. Latency and voice quality — test with your own scripts.
  3. Telephony support — numbers in your countries, SIP trunk import, call transfer, DTMF, voicemail detection.
  4. Tool / function calling and webhooks — can it call your CRM and calendar reliably?
  5. Analytics and recordings — transcripts, call logs, tagging, evals.
  6. Compliance features — recording controls, consent handling, data residency, PII redaction, enterprise agreements.
  7. Cost at your volume — all-in per minute including telephony, speech, LLM and voice. Model it at 1,000 and 10,000 minutes a month.
  8. Exit cost — can you export prompts, numbers and call data?

Our recommendation for your first agent

Use a hosted platform. Your goal is to learn conversation design, tool use and evaluation — not to debug audio streaming. Keep your prompt, knowledge base and tool definitions in your own files so you can move later. Revisit the stack once you have real call volume and a measured pain point.

Swap-ability rule

Keep three things in your repository, not only in a vendor dashboard: (1) the system prompt text, (2) tool schemas and webhook code, (3) your test cases. Those are your real assets.

Do this now

Pick two candidates. Score them on the eight criteria. Sign up for free trials and run the same 60-second script through both.

💡 Real-world example

Illustrative scenario. An agency built 12 client agents in a vendor dashboard with prompts only stored there. When pricing changed they spent two weeks re-creating them. The next 12 were built from a prompt library and tool schemas in version control — migration took an afternoon.

✅ Checklist

🛠 Your assignment

Write your final choice in one sentence with the top two reasons.

📝 Check your understanding

Answer 2 of 3 correctly to complete this lesson.

1. Which path is best for a first agent?
2. Which asset should you keep outside the vendor dashboard?
3. What cost number should you model?
Free lesson — create a free account to save progress, take the final exam and get your certificate. Next lesson →
💬 Questions about this lesson?
Ask the community — practitioners and the BFOUNT team answer.
Ask a question
Was this lesson helpful?