Where this is heading - and what stays human.
This is the final piece in our three-part series on how we've put AI agents to work inside AVV. The first covered what the agents do for us, while the second examined where they break. This one is about where we see this going - and what we intend to keep for ourselves.
Ask GP Eddie Thai where this is heading, and he goes beyond the org chart we have today. In his view of a year or two down the line, partners at AVV stop managing people as much and start managing systems. “Orchestrators of orchestrators of agent orchestrators," as he puts it, with "humans being the quality checker and the curator." Things happen quasi-automatically, while people sit at the top making decisions.
This may sound abstract until you learn that Eddie already runs a personal wellness agent that keeps his exercise routine and macro targets on track. The grand vision and the mundane reality arrived together, and he had waited for both: high-stakes work taught us to be cautious about relying on AI, and he held off applying it to his own life until the technology reached what he calls a "minimum viable" level of quality and care. Watching what Khoa and Ava, our main agents, can now do, and where they still fall short, is what finally cleared that bar.
So he's building a personal version of what the firm has: a Khoa-like orchestrator at the top, with narrower agents beneath it. What keeps it trustworthy is a short set of design principles carried over from AVV: don't fabricate context, speak up when something is missing, separate fact from inference, cite sources for anything material, and default to drafts that need human approval.
Partner Adrian Latortue expects this to pull in different directions for different people, since everyone's work and tools vary. A personal agent plugs into what you actually care about: the first things you read in the morning, prep for the day's meetings, a reminder that a new draft has landed or that someone never followed up. The part he's most interested in is agents talking to each other, such as his agent pinging a colleague's before a call, and even starting the call. "That becomes really collaborative," he says.
Partner Hau Ly, meanwhile, mostly wants the time to go deeper, learn the tools properly, and start building agents of her own.
There's a reason the same shape of an orchestrator over specialists keeps recurring. A general-purpose agent can't do each job as well as a focused one. Ava, for example, is most reliable at pulling from our databases precisely because that's all she does. The architecture ends up mirroring how the firm already works: a partner coordinating a bench of specialists.
One note of caution comes from our LP Capria Ventures, which is further down this path. Francis Perelman, their AI Program Manager, says his team now thinks hard before building anything from scratch, because the frontier models keep absorbing use cases. Something worth building six months ago may not be worth building today, when an off-the-shelf bot can do it without a workflow you'd have to maintain. The lesson for our own agent-building: build what's genuinely ours, and let the foundation models handle the rest.
Most of what our agents currently do is retrieval, while Adrian argues that the more interesting skill is rigor.
Today an agent fetches a number. Tomorrow it could assemble a complete picture of a company from data scattered across systems, adding layers of context no single person holds. From there it's a short step to agents that check our consistency while flagging where a deal or a hiring decision contradicts a process we've formalized. Or, they could undertake market intelligence that refreshes itself, so we're not, as he puts it, researching a sector from scratch every time a deal appears in it. Another possibility is valuation agents that combine the empirical (how a company's revenue is growing) with the qualitative (notes from calls that live in tools like Granola), and propose a mark for a human to accept or revise.
The point isn't automation for its own sake. Instead, it’s ensuring consistency that doesn't depend on how overloaded everyone happens to be. The agent gets you most of the way, and the human does the rest. "About 20 to 30% of the entire work," in Adrian's estimate, but the part that matters most.
Capria is already exploring the edges of this. They're testing agents that run around the clock in the background, reachable over Slack or WhatsApp, that you hand a task and let work. They're also considering training their own models on more than a decade of proprietary investment and company performance data. These aren’t fully implemented, but they paint a picture of what could be possible.
None of this means we will replace ourselves. Hau is direct about it: the agents won't take her job, "at least not yet," because using them well still demands real knowledge and judgment. "You don't trust the agents blindly," she says. "You have to have your own judgment." Eddie highlights the irreplaceable part in the qualitative gut, the X-factors of whether an opportunity makes sense. The human role shifts, but it doesn't shrink, and remains the final layer of quality on anything leaving the office.
There's an intriguing lesson within all of this. The best advice Hau has heard on adopting AI was to start small and focus on the problem, not the technology. Find work that is repetitive, manual, and unenjoyable, and begin there. As she points out, that's exactly what AVV tells founders, and how we approach investing. A firm that urges startups to solve real problems instead of pursuing fancy toys is taking its own advice.
So here's where we think we'll be in a year or two: not a firm that replaced its people with agents, but a network of human-curated agent systems, each of us orchestrating our own. We're sharing this as a work in progress since this story is far from complete, and we'd rather write it out in the open.
