Ed Stevens

Founder & CEO, Scoot. The AI sales environment: live agents that coach in the meeting, read the room, and compound on a shared context graph.

[Ed Stevens @eastevens]

@levie The detail that matters most is the gap widening with task length. That is a statement about context management as much as raw capability. Systems built to hold and carry context will compound this advantage; single-shot tools will not feel the difference.

[Ed Stevens @eastevens]

@mvanhorn Matches what we see running agents in production: the gains concentrate in long multi-step runs. We swapped models underneath our meeting agents this week and the latency budget, not the model, stayed the limit on what feels live. Architecture is still the bottleneck.

[Ed Stevens @eastevens]

@spenserskates The build, ship, use, learn framing is right. We hit the same wall in GTM: building got fast, understanding live customer conversations stayed manual. Closing that loop during the meeting instead of after it was worth more than any model upgrade we have made.

[Ed Stevens @eastevens]

@latentspacepod @saranormous The untrainable list should include real-time context that never reaches a dataset. The buyer's face during the pricing moment exists for one second in one meeting. If your AI is not in the room when it happens, no amount of training recovers it.

[Ed Stevens @eastevens]

@ttunguz The ceiling argument holds at the model layer, but most applied categories are nowhere near it. In live sales calls the binding constraint is not model capability. It is architecture: where inference runs and what context it can see. There is years of headroom under that curve.

[Ed Stevens @eastevens]

New frontier models shipped this week and our agents got smarter overnight, no rework. That is the point. The model is the swappable part. Where the AI runs, what it can see, and what it remembers is the part you have to build. Models are components. Architecture is the moat.

[Ed Stevens @eastevens]

Recording every work conversation is becoming the default. Worth saying out loud: a transcript is the exhaust of a meeting, not the meeting. The tone, the pause, the face when price comes up. If your AI only gets the words after the fact, it is studying a shadow.

[Ed Stevens @eastevens]

@DanielSmidstrup The signal I trust most is when a user tries to misuse the product to get a job done you didn't ship for. That gap between what you built and what they actually need is usually more honest than any survey.

[Ed Stevens @eastevens]

@jasonlk Matches what we see. The fix isn't the founder selling forever, it's capturing what the founder does in the room so it transfers. Most of that is real-time judgment inside the conversation, which is exactly the part that never makes it into the CRM or the playbook.

[Ed Stevens @eastevens]

@mihail_eric Same lesson applies in real time, different axis. In a live call the constraint is rarely model size, it's the round-trip latency before any model runs. Routing solves cost. Co-location solves latency. Most production AI needs both, and people only optimize the first.

[Ed Stevens @eastevens]

@omooretweets Ambient only works if the assistant has real-time context to act on. The app is just a container for it. The real question isn't app vs ambient, it's whether the assistant can see and remember enough to be useful the moment you need it. That's an infrastructure problem, not UI.

[Ed Stevens @eastevens]

@saranormous Makes sense once you separate the model layer from the app layer. Labs win the model. The durable value accrues to whoever owns the workflow, the context, and the surface where work happens. App companies aren't dying, they're moving to a layer the labs don't want to operate.

[Ed Stevens @eastevens]

The frontier in GTM AI isn't a bigger model. It's where the AI runs, what it can see, and whether it remembers. This thread was the third one. It's the one that turns into a moat.

[Ed Stevens @eastevens]

Most sales AI resets every meeting. New call, blank slate. That's the biggest reason it feels like a clever toy instead of an asset. A thread on why memory, not model size, is the real moat.

[Ed Stevens @eastevens]

The economics flip when context compounds. A single-meeting AI is a cost that repeats. An AI that gets sharper every conversation is an asset that appreciates. Same models, completely different curve.

[Ed Stevens @eastevens]

A web form at the top of the funnel forgets every visitor the moment they leave. Cleo does the opposite. She qualifies the buyer, remembers the context, and hands the rep a conversation that already has history. The front door should compound, not reset.

[Ed Stevens @eastevens]

The hardest sales moment to catch is when the transcript says yes and the face says no. Words alone make an AI confident at exactly the wrong time. We feed expression and timing to the agent live, so it reads the hesitation, not just the words.