The Truth Behind AI Agents For Investment Management
Reflections from Hypercore’s recent panel, featuring Hum Capital’s Blair Silverberg and Becca Gielowski
Last week, Hum Capital’s very own Blair Silverberg and Becca Gielowski joined Hypercore and Pinegrove Venture Partners for a panel on how AI agents are reshaping investment management. The conversation cut past the hype and got into the part that actually matters: where this technology earns its keep today, where it doesn’t, and what the next twelve months look like for firms willing to put it to work.
Here’s what stood out.
The unglamorous reality about agent architecture
LLMs are stateless. They don’t remember anything between calls. Every piece of “memory,” every reference to your last conversation, every awareness of what’s on your screen — all of it is engineered around the model, not inside it.
That framing matters because it explains why some agent products feel magical and others feel brittle. The magic isn’t in the model; it’s in the harness — the application layer that manages context, calls tools, summarizes when the window fills up, and decides what to load and when.
In evaluating agent platforms for high-context and high-scale use cases, the linchpin is understanding how they handle context, how they manage cost per request, and how they let you encode institutional knowledge.
Start where liability is low
The panel was unanimous on what not to automate yet: money movement, SEC compliance and archiving, and core engineering work. The reasoning is straightforward — these are the places where a wrong answer is expensive, irreversible, or both.
The first workflows worth trusting to an agent share a few traits. They have natural quality assurance layers downstream. They have human oversight built in. Or they’re creative tasks where the agent’s output is a starting point, not a final answer.
This is a useful filter for anyone thinking about their first deployment: where in your workflow does a mistake get caught quickly and with negligible overhead expense?
Trust earned through evaluation
The hardest part of deploying agents isn’t getting them working. It’s knowing when to trust them.
The panel converged on a few best practices: build golden datasets that encode what “good” looks like, benchmark agent output against expert human judgment, and ask agents to flag their own uncertainty before they act. As agent workflows ramp up, it’s important to know where your agent infrastructure accuracy bar is and work intelligently from that starting point.
Where this is heading
The panel touched on what the next year could look like for firms leaning into AI agents — fewer manual touches in core operating workflows, faster cycles for historically manual and cumbersome procedures, and richer feedback loops with key stakeholders.
The technology is already capable. What’s left is the hard work of cleaning data, encoding judgment, and building the evaluation infrastructure to know when an agent is ready to be deployed at scale.
