Where the Work Went Next

# Where the Work Went Next

A field report on what Waterr shipped since the last one — and the one insight per track we're carrying into the next build.

A voice agent on a live call now reasons about the whole conversation — what you said in minute four, what it promised in minute nine — and the reasoning costs zero added clock time. 22.5 seconds of wall time per conversation with it, 22.5 without. That result shipped this month as public research, numbers attached.

Three weeks ago I wrote down where the work was: a reasoning sidecar, a compiled-truth graph, a privacy layer, and a measurement discipline that killed a quarter of wrong building with one pre-registered experiment. That post ended with a claim — the frontier model is a rented pipe, and the moat is everything you build around it.

This is the follow-up. Three of those tracks moved. Each one produced an insight I didn't have in July, and the pattern across them surprised me: the same discipline that runs our research now runs our marketing.

1. The sidecar shipped as public research — Monologue

The sidecar has a name now, a public write-up, and its numbers attached: waterr.ai/research/monologue.

The short version. A realtime voice model gets 200–500ms to start speaking — that's the whole budget, and reasoning does not fit inside it. So the model forgets the constraint you stated in minute four and contradicts a commitment it made two turns ago. The industry's instinct is to wait for a bigger realtime model. We think that's the wrong bet — the model already knows what you said. It just doesn't have it in front of it at the moment it has to answer.

Monologue is a reasoning sidecar that runs beside the voice model instead of inside it. While you're still speaking, a text model reads the conversation and writes a short briefing — three typed slots, nothing free-form:

• MEMORY — what you actually said, quoted rather than paraphrased.
• CONSTRAINT — what the agent already committed to, and that it has to hold.
• TRAJECTORY — what the next reply needs to do.

That briefing gets injected between turns, into an audio gap that already exists. On AudioMC it scores 69.8% against 38.5% for the bare backbone. Self-coherence up 20.3 points. Inference memory up 16.3. Wall time per conversation: 22.5 seconds with it, 22.5 seconds without — the reasoning is free at the clock. It runs at $0.03 per conversation-minute with the planner included.

The insight: publish the receipt, not the adjective. The July version of this work lived in a blog post with a confidence interval. The August version is a research page where every number is checkable, and the difference in how people engage with it is not subtle. "Our AI is more coherent" is an adjective; +20.3 points of self-coherence at zero added wall time is a claim someone can attack — and a claim that can survive being attacked is the only kind worth publishing.

2. A year of deployments, two findings about trust

We've now put AI personas into real interviews, sales roleplays, discovery calls, and onboarding sessions across very different domains. I expected each domain to want different things. Mostly, they didn't. Two findings survived every deployment — and I heard both from customers before I saw them in our metrics.

Self-coherence is the trust threshold. One partner put it bluntly: the moment the agent forgot something it had said two minutes earlier, the whole session "felt off" — their words — and the participant checked out. Users forgive an AI for a lot. They do not forgive it for contradicting itself. Below that bar, nothing else about the product matters; above it, everything else gets a hearing. This is why Monologue's headline metric is self-coherence and not something flashier.

"Heard" is not "understood." Heard is transcription. Understood is adapting to intent — noticing the candidate is nervous and slowing down, taking a follow-up as a follow-up instead of restarting the topic. Watching session recordings taught me you can see the exact moment a participant decides the agent understands them: their answers get longer.

A voice agent is trusted when the participant stops performing for it and starts talking to it. You can find that moment on the recording's timeline.

The insight: the metric was in the recordings, not the dashboard. Answer length over session time is a trust signal no benchmark hands you — we found it by watching real sessions the way a coach watches game tape. If you build conversational AI and don't watch your recordings, you are optimizing numbers your users don't feel.

3. The go-to-market got rebuilt around the developer

In July we made a call: the customer we build the public surface for is a developer looking for an AI meetings API — not the end buyer two levels up. Everything public-facing got rebuilt around the person who will actually call the API. Guides written from the OpenAPI spec instead of from positioning documents. docs.waterr.ai promoted to the source of truth for every capability claim we make anywhere.

That last part became a rule with teeth: if the docs page doesn't exist, the capability doesn't exist — for marketing purposes. Not "it's in the codebase." Not "it ships next sprint." A capability we can't point a developer at is a capability we don't get to mention.

The insight: writing for the person with build intent fixes your writing for everyone. A developer evaluating an API punishes vagueness instantly — so every page had to lead with what the thing concretely does, what it returns, and what it costs to integrate. Those pages turned out to be better reading for non-developers too, because precision travels up the org chart in a way that marketing language never travels down it.

4. Comparison pages, verified against the other side's docs

The strangest work of the month: we rebuilt our competitive comparison surface, and the rule we set was that every claim about a competitor gets verified against that competitor's own live documentation on the day it's written — including honest "choose them if" sections where they're genuinely the better fit.

Two things happened that I didn't expect.

First, verifying competitors made them more impressive, not less. Reading a rival's docs closely surfaces real capabilities their own marketing undersells. The comparison got harder to win and more credible at the same time — which is the trade you want, because a comparison page nobody believes is a page working for the other side.

Second — and this is the part worth the whole exercise — the audit caught us first. Our own copy claimed a single API call returns the transcript, the recording, and the scores together. Our own docs say the recording is a separate endpoint, ready asynchronously, minutes later. The pre-publish audit flagged it and we rewrote the claim in fifteen-plus places before a single developer could hit the gap between the promise and the response body.

The insight: accuracy about your competitors is a trust strategy, and the same audit that keeps you honest about them keeps you honest about yourself. The overclaim you catch in review costs an edit. The one a developer catches in production costs the relationship.

The through-line

In July the through-line was that the model is a locked pipe and the compounding work is the harness around it — what the model gets to think about, what it gets to see, what is true, which builds get started at all.

The August lesson is that the harness posture doesn't stop at the product boundary:

• Monologue ships with its numbers attached, because a claim that can be attacked is the only kind that compounds.
• The trust findings came from watching recordings, because the metric users feel is not always the one the dashboard shows.
• The docs rule gates every capability claim, because marketing that outruns the docs is debt a developer will collect.
• The comparison audit verifies both sides, because the cheapest place to catch your own overclaim is before it ships.

Rented model, owned discipline. The centre is identical for everyone. The periphery — the harness around the model, and now the harness around what we say about the model — is the part that compounds every time we ship.

That's where the work went next.

*The previous field report is Where the Work Is Right Now. The load-bearing numbers live in the Monologue research note. The wider architectural argument is in The Voice Model Is Locked. We Built a Sidecar.*

*Try Waterr in a real meeting — interview, discovery call, training session — at waterr.ai.*

— Harshit