AI agents for your finance firm, set up for you

Every run saves the full request next to the brief, word for word, even when Claude is switched off, so your CCO can read it before turning the model on. The model gets these rules, the task, and market facts on the holdings list. No account numbers, no client files. After each draft, any number that isn't in the facts is listed in the audit log for the reviewer to check first.

SYSTEM
You write internal notes for staff at a registered investment adviser.
Rules:
- Use only the facts in the JSON provided. Do not add prices, news, or numbers that are not in it.
- Do not give buy, sell, or hold recommendations, and do not predict prices.
- If a fact is missing (for example no headlines), say so plainly instead of guessing.
- Plain English, short sentences, no hype. Markdown headings and bullets.

USER
Write this morning's brief for the advisors. Structure:
# Morning brief: <firm> (<date>)
## What matters today  (3-5 bullets, most client-relevant first; weight by clients_holding)
## Movers  (table: ticker, 1d %, 5d %, clients holding, why flagged)
## Earnings coming up
## Headlines on client holdings
## Data notes  (as-of dates, any tickers that failed to load)
End with the line: Internal use only. Not investment advice. Prepared by an AI agent from market data; review before sharing.

Facts (JSON):
{"firm": "Harbor Ridge Wealth", "date": "2026-10-04", "holdings": [{"ticker": "MSFT", "name": "Microsoft", "clients_holding": 41, "firm_weight_pct": 6.8, "asset_class": "equity", "as_of": "2026-10-02", "close": 517.530029296875, "prev_close": 512.7999877929688, "close_5d_ago": 516.1699829101562, "volume": 17803300, "avg_volume_20d": 21253995, "year_high": 553.719970703125, "year_low": 349.20001220703125, "next_earnings": "2026-10-28", "headlines": [{"title": "How Much Revenue Visibility Does Microsoft\u2019s Backlog Really Provide?", "publisher": "Insider Monkey", "date": "2026-10-03"}], "chg_1d_pct": 0.92, "chg_5d_pct": 0.26, "off_52w_high_pct": -6.54, "days_to_earnings": 24, "flags": []}, … 9 more holdings], "failed": []}

Every run writes append-only records: which inputs were read, what data was fetched, who ran it, and a hash of the output. If anyone edits a brief after the fact, the check flags it. Nothing is marked as sent until a person reviews it.

{"ts":"2026-10-04T08:56:57Z","workflow":"morning_brief","action":"inputs_loaded","operator":"r.patel","source":"data/holdings.csv","source_sha256":"3fe948ba…"}
{"ts":"2026-10-04T08:57:03Z","workflow":"morning_brief","action":"market_data_fetched","provider":"yahoo_finance","ok":10,"failed":[],"facts_sha256":"70cf8411…"}
{"ts":"2026-10-04T08:57:03Z","workflow":"morning_brief","action":"draft_generated","generator":"template","prompt_sha256":"672795f6…"}
{"ts":"2026-10-04T08:57:03Z","workflow":"morning_brief","action":"output_written","output_sha256":"5a3b9de4…","review_status":"pending_human_review"}

Vittoria: previously a senior software engineer at Amazon, now building AI agent systems at a tech company. Agent Desk is a small, hands-on side practice, so each pilot gets direct attention.

Yes. Export positions from your custodian or portfolio system as CSV (one file per account is fine) and the agent builds the holdings list from it: cash and sweep lines dropped, positions combined across accounts. The first brief can be on your own names the same day.

Into your own business AI accounts and tools, under your agreements with those vendors. We don't keep copies.

We design for it: logged actions, human approval before anything reaches a client, nothing that gives investment advice on its own. Your CCO has the final say, and we work with them.

Pilot setup is a fixed fee agreed on the scoping call, with an optional monthly care plan. Nothing is charged before you agree a scope.