Break, Optimise, Monitor: A Hands-On Voice Agent Workshop (Bangalore)
Your agent passed the demo. Then a real person called.
Agent: "Hi, I'm calling about your outstanding bal—"
Caller: "haan haan I know, listen, I already paid it on Tuesday" Agent: "I'm calling about your outstanding balance of ₹12,400. Would you like to pay now?"
Caller: "I just told you I paid." Agent: "I'm calling about your outstanding balance of ₹12,400..."
Eleven seconds, three failures. It didn't yield on the interrupt. It didn't retain what was just said. It looped. None of these show up in a demo, because in a demo you wait your turn, you speak clearly, and you never change your mind.
Anyone can ship a voice agent now. Almost nobody can tell you whether theirs is actually good. That gap is what we're working on.
What we'll do
You bring an agent, we try to break it, then you fix it and prove the fix held. Working session, laptops open, three hours.
1. Get an agent on the line (30 min) Connect what you already have (Vapi, Retell, LiveKit, or your own websocket / phone number), or start from the template we'll hand out. Chat agents welcome too, the same failure modes apply without the audio ones.
2. Build the callers who will break it. Generate a scenario suite from your agent's own description: personas crossed with situations. The impatient one. The one who trails off. The one who switches to Hindi mid-sentence. The one who asks the same thing four ways. Branching conversation graphs, not a flat script, so the caller reacts to what your agent actually says. Then we dial: real simulated calls against your live agent, running in parallel.
3. Score it the way a careful human would, on every call
One person can review 20 calls. Nobody reviews 1,000. So we score them automatically, on three layers.
Did it handle a human? Interruptions, loops, remembering what was said four turns ago, objections, escalating when asked for a person, asking instead of guessing, following a switch into Hindi, ending the call cleanly.
Did the voice layer hold? Transcription accuracy, speech accuracy, and the latency that decides whether you sound alive, broken out across transcriber, model and voice so you know which one to fix.
Would it get you in trouble? Leaked PII, prompt injection, invented policy, and tool calls that sounded right but did the wrong thing.
You'll see your own agent's scores, ranked worst first, and pick what to fix.
4. Fix it with optimiser agent and prove it Read the failure clusters, change the prompt or the flow, re-run the identical suite against the new version, and diff. Same scenarios, same personas, so the number means something. We'll also run the prompt optimiser against your eval set and see whether a machine beats your hand-written fix.
5. Keep it honest in production: Pre-production testing has an expiry date. The last block is the loop that keeps working after launch: tracing live calls with voice-specific latency and interruption data, clustering real failures instead of reading them one by one, alerts on eval scores rather than on crashes, and the part most teams never build, taking a real call that went wrong and replaying it back as a permanent regression test. Then wiring the suite into CI so a prompt change that breaks escalation never reaches a phone.
What you leave with
Your own agent, tested against a scenario suite you built in the room
A scored report showing where it actually breaks, not where you assumed
A before-and-after diff on a fix you made and verified
A regression suite and a CI gate you can run on Monday against the real thing
Who this is for
Engineers shipping voice or chat agents. Founders whose product is the agent. PMs who own a voice product and have to sign off that it's ready. Support, sales, collections, BFSI, healthcare, anyone whose agent talks to customers who don't behave.
Not a good use of your Saturday if you're looking for an intro to voice AI, a survey of TTS providers, or slides. You will be typing.
Want 10 minutes to show what you're building?
We're opening a few slots for people to talk about their own agent: what it does, what's working, and what's still breaking. Not a pitch slot. The room is engineers, founders and PMs building the same thing, so the useful version is the honest one, including the failure you haven't solved yet.
Open to anyone attending. To ask for a slot, email salil@futureagi.com or message Salil on LinkedIn with:
what you're building and who it's for
where it is today (prototype, pilot, live with real callers)
what you'd show in 15 minutes, or the problem you'd put to the room
Slots are limited and we'll confirm by email.
Bring
A laptop, and an agent if you have one (an assistant ID or endpoint is enough). If you don't, use ours. A free Future AGI account and the open-source repo are part of the setup in the first ten minutes, so get there on time.
Who's running it
Khushal Sonawat, Founding Engineer and Tech Lead at Future AGI.
"I break voice agents for a living, so they don't break in front of your customers. Anyone can make an agent talk. The hard part is knowing whether it holds up when a real person interrupts, trails off, or changes their mind mid-sentence. I stress-test conversational AI against thousands of simulated callers, personas, and edge cases before a single real user picks up the phone. What I care about is what it takes to trust a voice agent in production, and why 'it sounded fine in the demo' is where most teams go wrong."
Details
Saturday 22 August, 11:00 to 14:00. Future AGI office, 3rd Floor, Hanto Tranquil Centre, HSR Layout, Bengaluru. Free, capped so everyone gets hands-on help. Lunch after.