Architecture AI Updates: November 1, 2026

1. Arca Runs Ten-Day Agent Tasks on Per-Client Memory, a Custom Harness and Deterministic Guardrails

Arca, an AI-native wealth management firm, described its agent architecture in an October 31 Latent Space interview with lead AI engineer Aman Singh. The work is long-horizon rather than long-running: a single agent task can span up to ten days of stop-start progress with many actors, decision points and handoffs, so Arca built a per-client memory system that its agents write to, tracking what has happened, what is due in the next couple of days, what is being watched long term, and which person owns each action. A custom harness routes each task to a model chosen from internal benchmarks, with Claude as the primary provider alongside OpenAI models, and owning the harness instead of a managed one keeps the option to switch providers open. Agents are limited to analysis and proposals while trades and other money movement stay with human advisors, which Singh summed up as “agents prompt people more than people prompt agents”; he said all safety restrictions are fully deterministic so they take those decisions out of the agent’s hands, and Arca also runs checks against known attacks such as prompt injection. For data access, code an agent writes inside a sandbox is translated into a strict expected format, the original code is thrown away, and guardrails run on that format before it reaches the backend, which still lets agents aggregate and reshape client data to build a financial plan. A “meta-agent” reviews advisor feedback alongside agent traces and proposes changes to the harness. Source