RomanLogic voice agents · 19 Nov 2026
The voice brain sits between ElevenLabs and the model. It keeps a record of the call as it goes, checks every sentence before the caller hears it, and makes sure a lead reaches the owner even when the caller hangs up early. It has run on the Ambient Heating demo line since 18 Nov, with real handset calls every day.
Today an ElevenLabs agent sends each turn straight to its model, and whatever the model says is spoken. With the brain, ElevenLabs sends the turn to the brain instead. The brain answers with the same model, but it knows what the caller has already said and checks the reply before it goes out.
If the brain is down, ElevenLabs falls back to its own hosted model for that turn, so a call never goes silent. Switching an agent on or off is one command, and the agent's settings are saved before every change.
Before any real calls, we ran 36 simulated calls twice through the LDG Joinery prompt: once with the brain only passing the model's words through (what callers get today), and once with the brain checking. Claude played the caller, and every name, number and address was invented.
| Measure | Today's setup | With the brain |
|---|---|---|
| Calls where a usable number was given but no lead reached the owner | 5 of 36 | 0 of 36 |
| Leads with a wrong or missing name or number | 8 | 0 |
| Calls that asked for something the caller had already said | 5 | 1 (minor) |
| Time to the agent's first word, as the caller hears it | 1.10 s | 1.13 s |
Nothing scored worse with the brain on. The extra 0.03 seconds is not something a caller can hear.
Ambient Heating has been on the brain since 18 Nov. Each round of handset calls turned up something, and each fix went in with a test so it stays fixed.
Before
Now
The "after" also carries the caller's spoken correction from 12 to 21, which the old lead missed.
Where it stands
Ambient passes all 27 automated tests on the brain. Each automated run costs about 430 ElevenLabs credits, so the full set runs after every change.
The brain runs on Claude Haiku 4.5 through Rachel's Anthropic key, on a server that already runs the demo desks, so it adds no hosting cost. All figures are measured, not estimated.
| Item | Cost | Where the figure comes from |
|---|---|---|
| One real call, 1 to 1.5 minutes | $0.035 to $0.066 | Three Ambient handset calls on 19 Nov: every model call the brain made, including reading the caller's details |
| Per minute of call, roughly | about $0.04 | The same three calls |
| The model on ElevenLabs' side | $0.00 | ElevenLabs bills a call on the brain as "custom LLM" at no charge, so its own model charge for that call goes away; voice and minutes are billed as before |
| Building and testing all of it this month | $16.97 | The brain's whole Anthropic spend for November so far: simulated call runs, test packs and every live call |
| Monthly ceiling | $20 | A hard spend cap in the brain: it stops calling the model and lets ElevenLabs' own model take over before going past it (raised to $40 for November only) |
For scale: a client line taking 300 calls a month, at about a minute and a half each, comes to roughly $15 to $20 a month in brain model costs. Today the cap is one total across the brain. Giving each client line its own cap is a small addition, so a busy line could never run up an unexpected bill.
One-off cost to move an agent is time, not money: about two to three days for the first one (see below), plus around 430 ElevenLabs credits per automated test run while it's being tuned.
This would only happen if Roman wants it. Logistically it's straightforward: the switch itself takes minutes, and most of the work is teaching the brain that agent's own way of taking a call.
| Step | Effort | What it involves |
|---|---|---|
| Copy the agent | minutes | All work is on the copy; the live agent is untouched until the end |
| Write its flow | ½ to 1 day | What the agent collects, in what order, and how to spot each answer. The main piece of work |
| Match its lead | a few hours | Map the brain's lead onto that agent's own send tool and webhook, if they differ from the demos' |
| Point the copy at the brain | minutes | One setting on the agent, one line in the brain's config |
| Test and tune | about 1 day | The automated pack plus handset calls, the same way Ambient was done |
| Switch the live agent | minutes | One setting, with a one-command rollback |
So the first agent is about two to three days. Later ones are quicker, because most of the patterns carry over.
Figures from the 18 Nov simulated-call results, the Ambient handset calls of 18 and 19 Nov, and the brain's own spend ledger. Every name, number and address in the examples above was made up for the test calls.