AgentsAutonomy & tool use
Arca explains how its AI agents handle ten-day wealth-management cases while humans approve trades and client contact
Arca, an AI-native registered investment adviser that recently raised $64M, runs agents across roughly ten-day client cases using per-client memory, nightly reflection and verifier models, while humans keep sole authority over trades, money movement and client communication.
From the Latent Space podcast episode of October 31, 2026 and cited sources: Arca's agents are triggered by events. [1] [2]
SOURCES[1] Using Agents for Long-Horizon Financial Work — Aman Singh… latent.space[2] Transcript: Using Agents for Long-Horizon Financial Work —… rss.beehiiv.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 31, 2026 podcast episode published
- Covered
- Nov 2 edition
- Evidence
- 10 sources across 10 sites; transcript sourced
- 1Using Agents for Long-Horizon Financial Work — Aman Singh…latent.space
- 2Transcript: Using Agents for Long-Horizon Financial Work —…rss.beehiiv.com
- 3InvestmentNews: AI-enabled RIA Arca raises $64Minvestmentnews.com
- 4SD Times: Anthropic adds memory to Claude Managed Agentssdtimes.com
- 59to5Mac: Anthropic updates Claude Managed Agents with…9to5mac.com
- 6WealthManagement.com: AI-Native Wealth Manager Arca Raises…wealthmanagement.com
RisingHow it’s measured
YouTube video · 3.7K views · 32 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Arca's agents are triggered by events. A follow-up email goes to the agent already handling that thread. A new signal, such as a phone call, starts a fresh agent. That agent catches up by reading a per-client memory record of owners, next steps and long-term items. A daily reflection job keeps the record current. Agents carry out the analysis from start to finish and bring in the right staff member at the right point, but trades, liquidations and messages to clients always need a human. [1] [2]
Hard guardrails sit at the points where actions carry risk, not on the agent's reasoning. A fast language-model check reviews every database write. Incoming email, documents and calls are scanned for prompt injection. Outgoing data passes rule-based PII checks plus a vote among several small models. Finished artifacts are checked for things like raw account numbers or promised returns. Sandboxed code is translated into a strict query format before it reaches the backend. A change log records every agent and human action, and it serves both as an audit trail and as a trigger for agents. [1] [2]
Because Arca is the advisory firm, not a software vendor, it can require staff feedback. Every edit or cancellation comes with a reason and is paired with the agent's trace. Each night a meta-agent reviews these and suggests changes to the harness. An engineer still approves prompt changes, because agent-written fixes tend to bloat the prompt. Experts curate a file-based knowledge base that agents cite as they work. Decision models steer agents toward verified, current facts, for example to avoid using outdated tax brackets. [1] [2]
Arca keeps its own harness that works with any model. It chooses models with internal benchmarks, sending heavy work such as tax analysis to top-tier Claude or GPT models and background work to smaller ones. Anthropic's Managed Agents now offers similar memory and 'dreaming' features, but Arca declines to depend on it so it keeps control of its own intelligence layer. Anthropic's Schwab-backed suite for financial advisors adds competitive pressure. [1] [2] [4] [5] [8]
The interview gives a concrete blueprint for long-running agents in a regulated business: risk gates at the system's edges, audit logs that double as triggers, and staff feedback used as evals.
Read the full assessment
All performance details are Arca's own; it has published no error rates or time savings.
Executive brief
Arca says that in its wealth-management business, AI agents now ask human staff to act more often than staff ask the agents. Arca is a registered investment adviser (RIA). It raised $64M and manages more than $1B in client assets (InvestmentNews). Aman Singh leads its AI platform team. In a Latent Space interview, he describes agents working on client cases that last about ten days, using their own memory, verifier models and a nightly self-review process. Agents do not trade on their own, move money or contact clients without a human. Everything about how the system works comes from Arca itself. None of it has been checked independently.
What changed and event timeline
Anthropic adds memory to Managed Agents
Agents hosted on Anthropic's Managed Agents service can now keep context from one session to the next. Arca later chose not to use this ().
Arca buys Sandbox Financial Partners
Sandbox, based in Bethesda, managed about $682M. Granite Bay Wealth Management was also brought under Arca ().
Anthropic ships "dreaming," outcomes and multi-agent orchestration
"Dreaming" is a research preview in which agents review past sessions to improve their memory. Arca describes its own in-house process the same way ().
Arca exits stealth with $64M
That total is a $48.5M Series A led by General Catalyst plus a $15.5M seed round (;).
Anthropic launches Claude for Financial Advisors
Charles Schwab is the launch custodian. Other partners include BlackRock, Addepar, Envestnet and Orion ().
Latent Space publishes the Singh interview
It was recorded at AI Engineer New York 2026 and covers Arca's harness, guardrails and feedback loops ().
Capabilities and access
- Clients never use the AI directly. Every client gets a human advisor (InvestmentNews; transcript 01:51).
- Models: Claude is the main model, alongside OpenAI models. Arca previously used Gemini. Exact model versions are not disclosed (02:52–05:00).
- Agents can analyze a case from start to finish, but trades, money movement and messages to clients need a human (06:32).
Read the full section
- Clients never use the AI directly. Every client gets a human advisor (InvestmentNews; transcript 01:51).
- Models: Claude is the main model, alongside OpenAI models. Arca previously used Gemini. Heavy work such as tax analysis goes to an "Opus-tier," "Sonnet-tier" or large GPT model; background work goes to smaller models. Exact model versions are not disclosed (02:52–05:00).
- Agents can analyze a case from start to finish, but trades, money movement and messages to clients need a human (06:32).
- Arca takes no outside clients for its software. It is the advisory firm itself.
Technical analysis for researchers and developers
- Event-driven agents: a follow-up email goes to the agent already handling that thread. A daily reflection ("dreaming") job keeps that record current (07:42–09:15).
- Sandbox: the agent runs outside the sandbox and sends computation into it. Guardrails are applied to that format before anything reaches Arca's backend (16:27).
- Verification:
Read the full section
- Event-driven agents: a follow-up email goes to the agent already handling that thread. A new signal, such as a phone call, starts a fresh agent. That agent catches up from a per-client memory record of owners, next steps and long-term items. A daily reflection ("dreaming") job keeps that record current (07:42–09:15).
- Sandbox: the agent runs outside the sandbox and sends computation into it. Code the agent writes is translated into a strict query format and then thrown away. Guardrails are applied to that format before anything reaches Arca's backend (16:27).
- Verification:
- Every database write gets a quick check by a language model.
- Incoming email, documents and calls are scanned for prompt injection.
- Outgoing data passes rule-based PII checks plus a vote among several small models.
- Finished artifacts are checked for things like raw account numbers or promised returns (09:33–11:02, 17:41).
- Audit log: every change by an agent or a human records what changed, who changed it, where in the system and when. The same log also serves as a trigger for agents (11:18).
Claims and evidence
- Funding, assets and acquisitions: reported in the trade press (WealthManagement.com). Sources disagree on who led the seed round.
- A model that external benchmarks praised was too slow and costly on Arca's internal tests: this is an anecdote.
- AI compliance review beats human review: Singh says this, but offers no numbers (18:17).
Read the full section
- Funding, assets and acquisitions: reported in the trade press (WealthManagement.com). Sources disagree on who led the seed round. WealthManagement.com names Venrock alone; other coverage names Index Ventures and Venrock (Pulse 2.0).
- A model that external benchmarks praised was too slow and costly on Arca's internal tests: this is an anecdote. The internal benchmarks have not been published, and the transcript renders the model's name as "Luna" (03:28).
- AI compliance review beats human review: Singh says this, but offers no numbers (18:17).
- Agents found hidden fees and a non-fiduciary advisor for a prospective client: a single example with no outside confirmation (24:12).
- No reviewed source independently evaluates Arca's agents.
Context and prior work
- Arca's memory and reflection design looks a lot like Anthropic's Managed Agents memory and dreaming features.
- Anthropic has been moving into finance since Claude for Financial Services in July 2025.
- Rival firm Range plans to eliminate its advisor workforce as AI improves, the opposite of Arca's human-advisor model (InvestmentNews).
Read the full section
- Arca's memory and reflection design looks a lot like Anthropic's Managed Agents memory and dreaming features. Arca keeps its version in-house so it controls its own "intelligence" layer (14:05).
- Anthropic has been moving into finance since Claude for Financial Services in July 2025. Singh calls that move validating, but argues that a firm built around AI is different from a firm that bolts on AI tools (15:20).
- Rival firm Range plans to eliminate its advisor workforce as AI improves, the opposite of Arca's human-advisor model (InvestmentNews).
Limitations, safety and contested findings
- Arca has published no figures on error rates, verifier recall or how much time agents save.
- Models rely on stale facts with confidence, for example using 2025 tax brackets.
- Nothing in the reviewed sources describes how the system recovers when an action fails, because agents do not execute trades (20:02).
Read the full section
- Arca has published no figures on error rates, verifier recall or how much time agents save.
- Models rely on stale facts with confidence, for example using 2025 tax brackets. Arca counters this with "decision models" that push the agent toward a verified knowledge base or web search (18:51).
- Nothing in the reviewed sources describes how the system recovers when an action fails, because agents do not execute trades (20:02).
- Agents' own prompt edits tend to bloat ("every little fix becomes the whole prompt"), so an engineer approves each prompt change (21:48).
- Coverage of Arca's launch quotes only supporters (WealthManagement.com).
Business and practitioner implications
- Owning the operation helps: because Arca is the firm, it can standardize workflows and require staff to give feedback.
- Firm knowledge belongs in the product: Arca's knowledge base is a set of files that experts edit, and agents cite it as they work (22:47).
- Draw guardrail lines around risk: rule-based gates sit on money movement, client contact and data leaving the system, not on reasoning.
Read the full section
- Owning the operation helps: because Arca is the firm, it can standardize workflows and require staff to give feedback. Each edit or cancellation comes with a reason and is paired with the agent's trace. A nightly "meta agent" reviews these and suggests changes to the harness (20:43).
- Firm knowledge belongs in the product: Arca's knowledge base is a set of files that experts edit, and agents cite it as they work (22:47).
- Draw guardrail lines around risk: rule-based gates sit on money movement, client contact and data leaving the system, not on reasoning.
- Choose models with your own tests: keeping the harness model-agnostic and running internal benchmarks hedges against a single vendor. Anthropic's Schwab-backed advisor suite puts competitive pressure on Arca.
The source trail.
Sources (10)
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Transcript retrieved via published_transcript; language en. Timestamped text, not direct audiovisual review. Source: https://rss.beehiiv.com/podcasts/transcriptions/01a18ebb-53a8-7089-a56f-938e5771f6d5.vtt. Automatic captions/transcription may contain errors.
latent.spaceTranscript: Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
published_transcript
rss.beehiiv.com