Thanks to an invitation from Premji Invest, I spent a day at The Long Game Leadership Summit listening to operators from Brex, Navan, Hippocratic AI, Outreach, Writer, and others talk about what the next generation of enterprise AI actually looks like in production. This was not the pitch-deck version but a real debate from people who are shipping real systems on what works, what breaks, and what they'd do differently.

I left with sharper conviction about the technical bets we're making at Actively. Here's what stood out for me.

The winning architecture is constellations of narrow agents

This came up independently across fintech, healthcare, and travel, which makes the convergence hard to ignore. At Brex, Art Levy described their "Agent Mesh", an event-driven architecture where dozens of narrowly scoped agents collaborate probabilistically instead of following deterministic linear flows. They've re-architected roughly 80% of their codebase around this pattern, and the results were concrete: onboarding throughput tripled, fraud detection false positives dropped, and automation accuracy hit 70–99% depending on the task.

At Hippocratic AI, Munjal Shah runs a "constellation" of 31 models in parallel for every patient interaction, each small, fast, and specialized, so they can validate each other's outputs in real time. At Navan, Ilan Twig built a "zero critical hallucination" framework using a second LLM to monitor the first, layered over hundreds of specialized guardrail agents.

"Small, task-specific models often beat general-purpose giants when you scope the problem tightly and use bigger models as an oracle for edge cases." — Ilan Twig, Navan

These teams arrived at remarkably similar architectures from completely different starting points. That convergence tells you something fundamental about how reliable enterprise AI gets built.

At Actively, we're taking this pattern to its logical conclusion. When your customer base spans hundreds of thousands of accounts, the answer isn't one agent that tries to reason about all of them but it's one agent per account, each maintaining deep context about its specific domain, while sharing learnings laterally across the entire fleet.

Our per-account agents don't just work in isolation; they form a mesh of their own, surfacing patterns from one account that sharpen reasoning in another, while a layer of orchestration agents coordinates customer-facing actions. We wrote about this architectural philosophy in our post on building agents for enterprise, where OpenAI and LangChain builders landed on the same conclusion: scope agents tightly, and let them collaborate. The monolithic "do-everything" model is a dead end. The future is purpose-built constellations. What makes this architecture compelling is that the use cases aren't bolted on, they fall out of it naturally. Proactive pipeline generation, churn prediction, expansion signals, these aren't separate products, they're emergent behaviors of agents that are always on, deeply contextual, and collaborating across your entire book of business.

Context beats data and legacy systems can't provide it

May Habib, CEO of Writer, drew a distinction I think is under appreciated: the difference between data and context. Legacy SaaS platforms have mountains of data but lack the decision-making know-how, the reasoning traces that explain why a deal was won, why a customer churned. Without that layer, you can't build agents that make good decisions.

"Context beats data. Legacy SaaS has the data but lacks the know-how." — May Habib, Writer

This isn't just a summit talking point. It's becoming a full-blown industry thesis. Glean recently published a piece calling context the next data platform, arguing that AI agents can only automate real enterprise work if they can model the processes, relationships, and decision traces behind the data and not just the data itself. VC’s like Foundation Capital are calling context graphs AI's next trillion-dollar opportunity. And Gartner predicts agents will be embedded in 40% of enterprise applications by 2026; all of which will need some form of structured context to reason effectively.

At Actively, we believe in this deeply, especially for GTM where there is a lot of information fragmentation across many tools. Our bet is to aggregate the context among all of these fragmented systems around an account and keeping the objective stable through the lifecycle(s) of that account. This was not feasible in a pre-AI world and something that is critical and feasible today with a per-account agent that is context dense, long-running and singular minded towards the goal of progressing the account through the lifecycle.

The ROI only shows up when you reimagine the entire workflow and that requires top-down urgency

Amit Mitra from Outreach reinforced this from the revenue side. When marketing increases spend but there's no intelligent follow-up on resulting leads, the money is wasted. The ROI only appears when you reimagine the entire workflow end-to-end by letting agents own the sequencing, timing, and initial engagement rather than optimizing isolated steps. When Outreach's customers made that shift, they saw ~40% increases in new revenue per rep and up to 50% larger deal sizes.

"Stop dashboarding for humans. Start dashboarding the agents." — Amit Mitra, Outreach

But here's the part that doesn't get enough airtime: none of this happens without top-down urgency. Multiple panels surfaced the same pattern; first- and second-line leaders resist AI because they see it as a threat. The organizations capturing value are the ones where senior leadership is personally driving the change, not delegating it. The reframe that works: agents take over the low-value work (research, forecasting, follow-ups) so humans can focus on what they're actually good at: conversations and relationships.

That shift is starting to happen. At Actively we are seeing more executives are approaching AI not as a tool to optimize one workflow but as a foundation to rethink how their entire revenue organization operates. When you take that systems-level view, the narrative flips. AI stops being a threat to your team and becomes a force multiplier for what your team is already great at. Our vision for human-agent collaboration isn't about bolting AI onto one step of the funnel. It's about reimagining the entire motion: Agents own the research, sequencing, forecasting, and first contact, all of the high-volume, high-reasoning-load work that buries sellers today. This doesn't replace the human role, it elevates it. Sellers get to spend their time on conversations, relationships, and judgment calls, the things that actually move deals. Not a copilot for individual tasks but a collaborator across the full revenue lifecycle that makes every rep more effective.

Infinite minds, not incremental automation

One of my favorite take aways came from the healthcare panels when they introduced a framework that reframed how I think about what we're building: Co-pilot (human in the loop), Auto-pilot (autonomous execution), and what Hippocratic AI called "Infinite Pilot" - novel interventions that only become possible when you have AI operating at a scale no human team could match.

For example, Hippocratic AI called 1,700 patients who had ignored texts and letters about potentially cancerous lung nodules. The AI convinced 250 to schedule follow-up appointments — generating $2M in revenue from $2K in compute, while potentially catching cancer early. In another case, they called 16,000 vulnerable individuals during a NYC heatwave to conduct real-time heat stroke assessments. That intervention would have required 4,000 nurses.

It wasn't automating an existing workflow but it was creating an entirely new one that was previously impossible.

This reenforces the thinking Ivan Zhao at Notion wrote last year about Infinite Minds - where we are often building the future via the rearview mirror. But if we can see AI as the “steel” or the “steam engine”, we can now re-design our organizations around minds that never sleep.

At Actively, this abundance mindset shapes everything we build. Every account in your TAM gets its own dedicated, always-on agent that runs indefinitely, carrying forward everything it's seen, inferred, and learned about that account. These agents don't just sit in a chat sidebar waiting to be asked. They operate across the full hierarchy of human-agent interaction we've written about, from on-demand intelligence (pull) to proactive suggestions (push) to real-time ambient assistance to fully autonomous action. Each agent synthesizes data from CRM records, call recordings, emails, and product usage signals to build its own context graph, not just what happened on an account, but the decision traces that explain why it happened and what should come next.

The GTM motions that become possible at that scale, proactive outreach to dormant accounts, pattern-matching across thousands of renewal cycles, real-time competitive response, are the kinds of use cases that only emerge from this architecture. These aren't incremental improvements to existing workflows, they're entirely new motions that were previously impossible because no human team could operate at that breadth and depth simultaneously.

The next frontier isn't intelligence, it's EQ and persuasion

These "Infinite Pilot" interventions aren't just a scale story, they’re also a persuasion story.

"The next evolution of models isn't raw intelligence, it's EQ and persuasion." - Munjal Shah, Hippocratic AI

Those 250 patients who scheduled appointments had already ignored texts and letters. The agent succeeded because it could engage in a real conversation, handle resistance, and connect on a human level. That's a fundamentally different capability than pattern-matching or summarization, and it opens up categories of work that most people haven't even considered delegating to AI yet.

I think this is one of the more novel ideas from the summit, and it extends well beyond healthcare. As human-agent collaboration becomes the default mode of working, not just for patient outreach but for sales, support, engineering, and everything in between, raw intelligence becomes table stakes. What separates a good agent from a great one is the same thing that separates a good coworker from a great one: not just being smart, but being someone you actually want to collaborate with day-to-day. High EQ, good judgment about when to push and when to listen, awareness of context beyond the immediate task. We don't just want brilliant coworkers but we want ones we enjoy working alongside.

At Actively, we have been spending more time as we design agents that work with sellers throughout their entire day; researching accounts, prepping for calls, drafting follow-ups, flagging risks. When an agent is deeply embedded in the every day workflows, EQ stops being a nice-to-have and becomes essential. It needs to know when to surface an insight versus when to stay out of the way, when a deal needs urgency versus patience, and how to adapt to the way each seller actually works. The closer agents get to being true collaborators, the more the bar shifts from raw intelligence to judgment, timing, and trust.

Closing Thoughts

Across every panel, fintech, healthcare, security, hardware, autonomous vehicles, the same patterns kept surfacing. Multi-agent constellations with narrow scopes. Context as the critical layer above raw data. An abundance mindset that reimagines what's possible rather than just automating what exists.

What struck me most is that nobody was debating whether agents work. The conversation has moved to how: what architectures hold up, what breaks at scale, and where value accrues. These are the exact problems we're working on at Actively as we build the next generation of enterprise agents.

If you're an engineer and you love tackling things like multi-agent architectures, context graphs, and long-running agents that learn and adapt, come build with us.