Aman Singh, who leads AI at financial-services firm Arca, describes its agent-driven workflow in this 13-minute AI Engineer talk.

A client goal can last days

  • Raising cash for a tax bill involves advice, asset sales, settlement and a bank transfer, with people responsible for different steps.
  • The agent keeps track of that goal across pauses and handoffs; it does not need to generate tokens continuously for days.
  • Singh says Arca operates the services business itself, giving it control over both the software and the working process.

Signals keep the work moving

  • Background agents process emails, calls, documents, custodian events and market changes. A custodian holds and administers clients’ assets.
  • A working agent can register interest in a reply or transfer completion, then resume when the relevant signal arrives.
  • Scheduled wake-ups cover missing signals. Staff-entered notes, calendar changes and completed tasks also generate events.

Memory and tools prepare the next step

  • Operational context tells agents how the firm handles a situation; client memory preserves preferences from earlier conversations.
  • In Singh’s example, the agent remembers a request to leave one stock holding untouched and revises the plan when the cash need changes.
  • The harness, the software running the agent’s tools and checks, includes a sandbox for financial calculations and persistent work state.

People approve and execute consequential steps

  • Agents create shared spaces containing plans, discussion and requests for the right colleague’s involvement.
  • In the tax example, people approve the plan and outgoing email, and a trader is brought in to handle asset sales.
  • Work that can be done now becomes a task; future obligations stay in agent memory until they become actionable.

Coordination needs its own engineering

  • Singh describes six months spent closing context gaps, improving task routing and maintaining state across interruptions.
  • Arca also standardizes its own business processes and turns colleagues’ explanations into evals, tests of agent behavior.
  • His proposed proactive workflow prepares clients’ tax situations each January and brings plans to staff for approval before clients ask.

Caveat: this is Arca’s own operating account. The talk provides no measured reliability or time-saving results, and human approvals remain part of the workflow.

“Agents prompt people more than people are prompting agents.” — Aman Singh, 11:00