Building a No-Code Multi-Agent Platform PMs Actually Use
Notes from shipping Justdial's internal voicebot builder from concept to 25,000+ calls/day in a month.
The hard part of a no-code agent platform isn't the agent framework. It's convincing a PM with no engineering background to trust it with a live vendor call queue.
The constraint that shaped everything
We had one month to take Justdial's multi-agent voicebot platform from concept to production, adopted by internal teams for daily vendor outreach. That constraint ruled out building a general-purpose agent framework and forced a narrower question: what's the smallest set of building blocks a PM needs to compose a working voicebot without writing code?
The answer was pluggable LLM/STT/TTS options behind a simple builder UI (Next.js + shadcn UI), backed by an in-house dialer and LiveKit for the call layer. PMs pick models the way you'd pick a font — good defaults, easy to override, hard to break.
Why pluggable, not fixed
Locking the platform to one LLM or one voice vendor would have shipped faster, but it trades away exactly what internal teams needed: the ability to launch scrappy and swap in better models later without rebuilding the flow. Pluggability turned out to be the feature that let usage scale to 25,000+ calls/day, because teams could tune cost and quality per use case instead of accepting one global tradeoff.
What made PMs trust it
Evaluation and monitoring, not the builder itself. A PM will forgive a clunky flow-builder UI. They won't forgive shipping a voicebot they can't see failing in real time. Giving every team visibility into call outcomes, not just call volume, is what turned this from a novelty into a tool Operations relied on for daily targets.