For Ed Stevens, Founder & CEO at Scoot, building AI agents is not about creating one giant system that tries to do everything. It is about structuring work in a way that is easier to build, debug, and improve over time.
That is why his perspective is so practical. Instead of treating an agent as a monolithic workflow, Ed describes a model in which teams “break these down into agents and sub-agents.”
Ed’s framework comes through in four clear ideas:
1️⃣ Start with one centralized GTM brain
Ed’s approach depends on orchestration. Individual agents may each handle a specific task, but something has to connect and coordinate them. Rather than operating as isolated workflows, they need a centralized brain that can direct the work, manage dependencies, and keep everything moving.
It becomes a hierarchy, where workflows can “sort of nest,” call one another, and be guided by a supervisory layer that brings the system together.
2️⃣ Focus on purpose-built agents with one job
Ed’s point is simple: smaller pieces make better systems. When teams try to pack everything into one workflow, it becomes harder to manage. As he puts it, if there were “a hundred of these nodes, it’s really hard to debug something.”
So the better approach is to “break them down into smaller pieces” that can be built in “more logical chunks.” That makes them easier to troubleshoot, easier to reuse, and easier to adapt from “customer to customer and use case to use case.”
3️⃣ Select the right LLM or tool for the job
Ed also makes clear that agent design should stay flexible. The point is not to force every workflow through one model. It is to match the model and tool to the task.
His framing is practical: if a client already has “an enterprise relationship with Anthropic,” then “you would just use Anthropic there.” The model choice should follow the job and the customer environment.
4️⃣ Keep a human in the loop
Ed is optimistic about where agents are going, but he is not pretending they run perfectly on their own today. He says Claude will get you there “97 times out of 100,” but the other “3 times out of 100” can become “a catastrophe” if no human understands what is happening.
Where this comes alive: real-time AI coaching inside the meeting
The clearest test of this architecture is whether it can actually perform under live conditions — when a salesperson is on a call with a buyer and needs help right now.
That is exactly the workload Scoot built The Brain for. The Brain is Scoot’s central agentic intelligence platform — the supervisory layer Ed describes above, plus the shared context store that every agent reads from and writes back to. It connects CRM data, conversation history, deal context, and product knowledge into one place, then routes work to the right purpose-built agent at the right moment.
Inside Scoot Engage — Scoot’s meeting and webinar platform — The Brain powers Live Advisor, a real-time sales coach that listens to the conversation as it happens and prompts the seller with the next best move. Because Scoot Engage runs its own video infrastructure (the Wizard SFU), the live AI agents are co-located with the meeting itself. There is no round-trip out to a third-party platform, no waiting on an external API, no latency penalty. Coaching shows up in the seller’s screen the moment it is useful.
A few things become possible once The Brain is the connective tissue:
Coaching that knows the deal. Live Advisor doesn’t just react to what is said in the meeting — it pulls the account history, prior call notes, and pipeline stage from The Brain, so its prompts are specific to this buyer, not generic best practices.
Insights that compound. What Live Advisor learns in the meeting flows back into The Brain. The follow-up agent, the proposal agent, and the forecasting agent all start from a richer picture for the next interaction.
Smaller, swappable agents. Live Advisor itself is a purpose-built agent — exactly the “one job” pattern Ed described. It can be improved, model-switched, or replaced without rewiring the rest of the system.
This is the framework Ed laid out, running in production. Smaller agents. One Brain. The right model for each job. A human still in the seat, now better supported.
In conclusion
Ed Stevens’s perspective on AI agents is grounded in real implementation: build smaller agents with a clear job, choose the right model for the task, use a master system to coordinate them, and keep people involved as the system is tuned. That is what turns an agent from a demo into something a team can actually trust in production.
About Scoot
Scoot is the AI Sales Environment — the place where AI agents, buyers, and sellers interact and get business done. Scoot is built on three pillars: Superior Meetings (Scoot Engage, meeting and webinar platform with Mingle, Presentation, and Meeting modes), Superhuman Agents (live AI agents like Live Advisor plus custom agents built for each customer), and Supreme Control (clients own and control their AI). Because live agents run co-located with Scoot’s own video engine, they are faster, cheaper, and more performant than approaches that have to round-trip data out of a third-party meeting platform.
Learn more at scoot.app or book a demo.
About Ed Stevens
Ed Stevens is the Founder & CEO of Scoot. A serial entrepreneur (Stanford → Shopatron → Kibo → Scoot), he is building the next generation of go-to-market software: a platform where humans and AI agents collaborate live in meetings to move deals forward.
