Inside Arca's Long-Horizon Agents That Run the Business
B · Useful, limited Arca's AI lead gives a clear first-hand account of an event-driven agent design inside a wealth-management firm; the architecture is concrete, but the talk carries no accuracy, volume or cost figures and ends as a recruiting pitch.
Summary based on the full transcript·Original: about 13 minutes to watch
In short
A 13-minute AI Engineer talk (uploaded 23 October 2026) by Aman Singh, who leads AI at Arca, a firm that sells financial planning, tax planning and investment advice. Arca is the service firm, not a software vendor. Its work is "long horizon, not long running": one client goal spans about 10 days of stop-start steps, hand-offs and settlement waits. Agents watch signals, keep per-client memory, publish plans for people to approve, and create tasks that pull in the trader or operations staff when their step comes.
Key numbers
- 10 daystypical span of one client goal, such as raising cash for a tax bill
- 8 monthsage of a client instruction the agents still recall from per-client memory
- 6 monthstime spent closing context gaps, task precision and long-horizon state
- 90%share of work the speaker estimates agents have done before a person is pulled in
As reported by the source.
English·Video·opens youtube.com
What it reports
- Worked example: a client owes $100,000 in tax. An advisor works out the tax-optimal sale, a trader liquidates, operations trigger the bank transfer; settlement and transfer each take days.
- Agent inputs: emails, calls, documents, meetings, custodian events and market moves, read continuously by background agents.
- Harness: tools, guardrails, a sandbox for financial computation and persistent per-client memory ("never touch Jane's Apple position", said on a call 8 months earlier). The speaker says harnesses are converging on a standard shape and the harder work sits outside them.
- Agents publish a plan in a shared "space" where several people can chat with the agent and approve it. A plan cannot execute without people.
- Agents register interest in events ("tell me when Jane replies", "when the transfer lands") and wake on a schedule if the event never comes. Human actions in the interface, such as rescheduling or marking a task done, also wake agents.
- Agents notify each other, for example when the cash need drops from $100,000 to $80,000.
- Over 6 months the team worked on context gaps, task precision and holding state. Because Arca is the firm, it standardised how work is done and turned that into agent context; staff reasoning collected in the product becomes evals.
- Staff log in to task lists where agents have done "90%" of the work, in the speaker's estimate. "Agents prompt people more than people prompt agents."
- New business goals, such as preparing every client's tax position each January and sending one-click approval emails, are set by changing the agents' goals.
How they got it
Conference talk with architecture slides; the transcript is from auto-generated English captions. No measurement is presented.
Limits
- No figures for accuracy, error rates, client numbers, cost or time saved; the 90% is an estimate.
- No description of guardrail failures or how wrong plans are caught beyond human approval.
- The talk closes with a hiring call; the firm has reason to present the system well.
- A related interview with the same speaker was not reviewed here.
Why it matters for companies
Most enterprise work looks like Arca's example: days of waiting, many hand-offs, small decisions at each step. The talk describes a design for that shape: agents subscribe to events and wake on them, people approve plans and execute the regulated steps, and every user action is itself an event. The precondition is the one firms skip: Arca first standardised how its own work is done. Without that operating context, the agents have nothing to follow.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com