{ "video": { "id": "oP2WLFpMoDY", "title": "Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca", "duration": 796, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "oP2WLFpMoDY", "title": "Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca", "one_liner": "Arca runs its financial-services firm on long-horizon agents. A signal system lets agents watch external and internal events, operational context tells them what to do next, a harness with persistent memory does the work, and a shared space pulls in the right person at the right time. The result is that \"agents prompt people more than people are prompting agents.\"", "summary": "Aman Singh, who leads AI at Arca, says that being AI native means agents run the business and execute the decisions that shape client outcomes, rather than assisting from \"some co-pilot corner.\" Financial-services work is \"more long horizon and less long running\": a single client goal, such as raising cash for Jane's tax bill, can stretch over 10 stop-start days with many actors, decision points and hand-offs. Arca hands that whole outcome to agents end to end, and the talk follows one client email through signals, plan, approval, waiting, replanning and human hand-offs. He ends by arguing that you know you are AI native when you can steer your business by steering your agents.", "key_points": [ "Arca is not a SaaS company. It calls itself \"an applied AI company that happens to be running a services business\" (financial planning, tax planning, asset management, investment advice), which lets it change and standardize how the firm itself works.", "Example workflow: client Jane owes $100,000 in taxes by April 15. An advisor analyzes statements to find the most tax-optimal way to raise the cash, a trader liquidates assets, settlement takes days, then ops triggers an ACH that also takes days. Across many clients, the time this takes \"exponentiates.\"", "Background agents continuously evaluate signals: emails, phone calls, incoming documents, meetings, custodian events and market movement. Agents use tools to register interest (\"let me know when Jane responds\", \"let me know when that ACH lands\"), the signal system routes events to the agents that asked, and agents also wake themselves on a schedule in case a signal never arrives.", "Actions inside Arca's own UI are signals too: adding information from an in-person meeting, rescheduling a calendar event, marking a task complete. Agents also notify each other; in the example, new information changes the plan from $100,000 to $80,000.", "The harness includes tools, guardrails, persistent memory over long horizons, deep client memory (e.g. never touch Jane's Apple position because she said so in a call 8 months ago) and a sandbox for financial computation. Aman says harnesses are \"coalescing on a standard shape\" and that what happens outside the harness is more interesting.", "Agents build a shared \"space\" holding the plan. Several people can chat with the agent there, and it is where agents pull people in to approve the plan, approve emails or execute steps such as trading, based on the operational context.", "Their framing of the primitives: stay aware (signals), know what to do next (operational context), do the work and keep track of it (harness and memory), bring in the right people at the right time (tasks). Work for the future is tracked in agent memory; work doable now becomes a task.", "Over the last 6 months they have focused on closing context gaps, task precision, holding state over long horizons, and operationalizing firm processes into agent context. Because colleagues use the product, their rich reasonings become evals. When users log in they see tasks agents have prepared after doing about 90% of the work. In the proactivity example, adding a new goal (\"every January, prepare each client's tax situation and proactively reach out\") makes agents produce analyses and one-click approval emails with full plans for every client." ], "takeaways": [ "For multi-day, multi-actor workflows, build a signal system where agents subscribe to events (external events, user actions in your own UI, messages from other agents) and also wake on a schedule as a fallback, so no human has to babysit the goal.", "Split long-horizon state explicitly: work for the future lives in agent memory, and work that can be done now becomes a precisely targeted human task.", "Treat the harness as a commodity and invest in what surrounds it: signals, operational context, and how and when humans get pulled in through a shared plan space.", "Standardize your own processes and encode them as agent operational context, and capture the reasoning of the people using the system as evals.", "Aim for agents that prompt people and drive business goals: adding a goal to the agents' goal set should be how you steer the business." ], "topics": [ "agents", "long-horizon agents", "event-driven architecture", "agent memory", "human-in-the-loop", "evals", "fintech", "ai-native" ], "tools": [ "Arca" ], "quotes": [ { "text": "In financial services, the work is more long horizon and less long running.", "at": "01:25", "url": "https://www.youtube.com/watch?v=oP2WLFpMoDY&t=85s" }, { "text": "So we let our agents watch the world. So people don't have to, right?", "at": "06:16", "url": "https://www.youtube.com/watch?v=oP2WLFpMoDY&t=376s" }, { "text": "agents prompt people more than people are prompting agents", "at": "11:00", "url": "https://www.youtube.com/watch?v=oP2WLFpMoDY&t=660s" }, { "text": "In an AI native company, you know you've built one when you can steer your business by steering your agents.", "at": "12:42", "url": "https://www.youtube.com/watch?v=oP2WLFpMoDY&t=762s" } ], "words": 2542 }, "summary_url": "/#oP2WLFpMoDY", "transcript": { "html": "/transcripts/oP2WLFpMoDY.html", "txt": "/transcripts/oP2WLFpMoDY.txt", "vtt": "/transcripts/oP2WLFpMoDY.vtt" } }