Latent Space

Aman Singh on running a firm where agents prompt people

Aman Singh· Leads the AI platform team at Arca at Arca
·~26 min·English·Latent Space
AgentsAI SafetyAI InfrastructureBusiness Strategy
TL;DR

Aman Singh, who leads AI at the AI-native wealth firm Arca, explains how owning the harness, persistent client memory, deterministic guardrails, and edge-based injection defenses let agents run proactive, multi-day financial workflows with a human in the loop for anything that moves money.

01Business Model

We're not a SaaS company. We're the firm.

<strong>Arca isn't selling agent software; it's the wealth-management firm itself</strong> — which lets it standardize how the work happens so agents can run it.

We're actually not a SaaS company. We are we are the firm ourselves.

— Aman Singh, Latent Space
Key Insight
Being the firm, not a vendor, flips the usual SaaS constraint: Arca can redesign the underlying workflow instead of bending to each customer's. The moat isn't the model — it's the standardized, owned process the agents run inside.
02The Inversion

Agents prompt the people, not the reverse

<strong>At Arca the agents prompt the people</strong>, running proactively 24/7 and pulling in a human advisor only at the step that needs one.

Agents prompt people more than people prompt agents and it is a proactive kind of service that is running for them 24/7

— Aman Singh, Latent Space
Key Insight
The inversion only holds because the service is proactive: if clients had to prompt, the system would just be a tool again. Keeping AI invisible to the client is a product decision that leaves the human advisor as the trusted face while agents do the back-office labor.
03Long-Horizon State

Work spans days; a fresh agent wakes to each signal

Servicing a client's tax bill can span 10 days, so <strong>a persistent memory system lets a fresh agent wake up to each new signal</strong> and resume where the work left off.

there's like a fresh agent that wakes up to figure out what this means

— Aman Singh, Latent Space
Key Insight
Long-horizon agents don't need one long-lived process; they need durable shared state. Arca's memory-plus-nightly-reflection is an external working memory any short-lived agent can rebuild context from — the pattern that makes multi-day agent work tractable at all.
04Guardrails

Guardrails sit at the edges, and they're deterministic

<strong>Safety restrictions are fully deterministic and take the decision off the agent</strong> — anything that moves money, liquidates assets, or contacts a client stops at a human.

we have strict guardrails around human in the loop for things that would either move money or liquidate assets

— Aman Singh, Latent Space
Key Insight
Deterministic guardrails are a deliberate choice to not trust the model on the highest-stakes actions. By containing autonomy at the edges of capability, Arca lets agents be fully autonomous on low-stakes analysis while making the irreversible, money-moving steps un-automatable by design.
05The Lethal Trifecta

Treat every email, doc, and call as an injection target

Emails, documents, and phone calls are all prompt-injection targets, so <strong>Arca verifies at the edges — scanning what comes in and checking what goes out for PII and exfiltration</strong>.

there's emails coming in, documents coming in, phone calls being made, which are targets for prompt injection

— Aman Singh, Latent Space
Key Insight
Arca maps its defenses onto the two places untrusted data crosses a trust boundary — ingestion and egress — which is exactly the lethal-trifecta failure mode of untrusted input, private data, and an external action. The consensus of small models shows the defense can be cheap enough to run on every crossing.
06Owning the Stack

Own the harness; your benchmarks beat the leaderboard

<strong>Arca owns a model-agnostic harness and lets its own internal benchmarks, not public leaderboards, guide model choice</strong>, swapping out any model that proves too slow or too costly in practice.

despite the external benchmarks claiming great things for Luna, it didn't really quite work out for us

— Aman Singh, Latent Space
Key Insight
Owning the harness is the same bet as being the firm: whoever owns the memory and operational knowledge owns the ability to shape the business. Outsourcing that to a managed-agent product would hand a vendor the one thing Arca treats as its edge.
07Grounding

Models will confidently use the wrong year's tax brackets

Finance models answer with confident but stale facts, so <strong>a decision model detects when financial rigor is needed and nudges the model to a verified knowledge base</strong>.

models will very happily use the tax brackets from 2025

— Aman Singh, Latent Space
Key Insight
The failure isn't that the model lacks knowledge — it's that it is confidently wrong, which is harder to catch than an obvious error. A decision model acts as a routing layer that recognizes when a question demands ground truth and forces retrieval, turning silent confidence into a checkable fact.
08Production Evals

Product feedback becomes evals, but a human gates the change

Product feedback is paired with the agent's trace and a nightly meta-agent proposes harness changes, but <strong>a human engineer is still in the loop before any prompt change ships</strong> because agents overfit each fix into the whole prompt.

every little fix becomes the whole prompt

— Aman Singh, Latent Space
Key Insight
Arca's self-improvement loop keeps a human at the step that matters most. The observation that agents can't stay meta is a sharp limit on automated prompt optimization: agents fix the specific failure in front of them, not the general policy, so generalization still needs a person.