Arca's agents run the business. Most teams still assist humans.

The claim. Arca, an AI-native financial services firm, has built an agent system that runs the business rather than assists it. Aman Singh describes a multi-agent architecture where background agents read emails, monitor market signals, and execute multi-day financial workflows with minimal human intervention. Humans are pulled in only for approval or execution, effectively inverting the typical copilot pattern.

How it works. A client email triggers a chain of agent activity: understanding the request, consulting client memory, running financial computations, and producing a plan. Agents watch for responses, market changes, and internal system events. They can register interest in future signals, wake themselves on schedules, and pull in specific colleagues at decision points. The system treats every interaction—task completion, calendar changes, document uploads—as a signal.

The operational difference. The key structural change is removing humans from the critical path. Instead of people monitoring progress over days, agents hold persistent memory, track state, and escalate only when judgment or execution is needed. Singh reports that agents prompt people more than people prompt agents. This is not a demo of a tool; it is how their firm operates today.

The trade-off. This approach requires deep standardization of business processes. Arca had to codify operational context, define handoffs, and build a signal system that routes events to interested agents. The complexity is emergent: hundreds of simultaneous signals, agents communicating with each other, and tasks appearing on human dashboards. That operational burden is invisible in the talk.

What is missing. The talk provides no failure data, error rates, or edge case handling. There is no discussion of what happens when agents make incorrect decisions, how human override works in practice, or how compliance and audit are handled in a regulated industry. The claim that a new business goal can be added by simply steering agents is presented without evidence.

The takeaway. This is a strong production-experience talk from a firm that actually runs agents in customer-facing work. The architecture details are useful for teams designing long-horizon agent systems. But the hype about steering businesses by steering agents needs substantial proof before it becomes a general pattern.