How Our Daily Standout Research Agent Works

Every weekday at 10:20 in New York, about an hour after the US market opens, a research agent built on Claude Managed Agents starts its run for moneyheap. It scans several hundred tickers, keeps the few large US companies rising on fresh, confirmed company news, researches each one, and files a shortlist with one standout and a handful of alternatives, plus the day’s biggest movers. The Daily Edition page, each day’s public page (for example 5 November 2026), the dashboard card and the morning email all read what it writes.

Why did we move the Daily Standout to an agent?

Because a fixed pipeline could only rank numbers, and the part readers cared about was the reasoning. An agent can read the news behind a move, check it against a second source, drop duplicate stories and write down what it couldn’t verify.

The first version ran inside our own backend as a fixed sequence: take the day’s top gainers, filter out the illiquid ones, run moneyheap’s standard analysis on a few, and have a model choose one. It worked, but it had two blind spots. It only saw the gainers list as numbers, so it couldn’t tell a rise on an earnings beat from a rise on no news, and it never looked at the decliners or at what people were discussing. And every step was hard-coded, so improving the research meant a code change and a deploy.

The agent version fills the same analysis shape the app already used and replaces the fixed steps with a research brief: what to scan, how to filter, what evidence counts, and what to write. It starts from a wider universe, checks the news behind each move, and reports the biggest decliners too, in a separate movers list. Changing how the research works is now an edit to that brief, not a release.

Where does the agent run?

It runs on Claude Managed Agents, Anthropic’s hosted service for long-running agents, not on our servers. A scheduled deployment starts a new session every weekday morning, and the agent’s system prompt is the source of truth for how it researches.

Three pieces make this work:

• The agent: a model, a system prompt and the tools it may use (web research, market data, and writing to our database). Each change to the prompt creates a new numbered version.
• The scheduled deployment: runs the agent on a cron schedule in New York time. It is pinned to one agent version, so editing the prompt doesn’t silently change tomorrow’s run. We update the prompt, read the new version, then move the deployment to it on purpose.
• The environment: the sandbox the session runs in, with network access to the data sources it needs.

There are two copies of the agent. The live copy writes the day’s research for moneyheap.ai and then copies the same records to our test site, so both show one run’s research at one run’s cost. The test copy writes only to the test site and runs on demand: a prompt change is tried there first and reaches the live copy only after review.

Keeping the agent outside our codebase was a deliberate choice. The app and the agent share only a data contract, so either side can change without the other noticing, as long as the contract holds.

How does the agent narrow the market down?

It works as a funnel: a wide universe of tickers, a size and direction filter, a fresh-catalyst filter, quality checks, then ranking. Each run writes the funnel’s counts and a plain-language account of every step, and the Daily Edition page shows both under “How we picked”.

Here is the funnel from the run on 5 November 2026, as the agent recorded it:

Step	Tickers left	What it kept
Universe	405	the 200 most-discussed stocks and funds on social media, the day’s 100 biggest gainers, 100 biggest decliners and 100 most active stocks, and every company named in the past day’s market news, plus sector and benchmark funds for context
Size and direction	32	US-listed companies worth $10 billion or more, rising 3% or more about an hour into the session
Fresh catalyst	14	rises explained by company news from the last day (earnings beats with raised outlooks, large capital-return plans), confirmed by at least two independent sources
Quality checks	9	set aside names with no confirmed company catalyst and a tracking stock
Shortlist	7	one standout and six alternatives

The cut-offs are part of the brief, and we still tune them: on 4 November the agent scanned 1,289 tickers with a $5 billion and 4% cut. Falling stocks now go to the biggest movers, not the shortlist.

The catalyst filter does more work than it looks. On 5 November, a sharp down day for tech stocks, many companies rose only because money rotated into defensive sectors such as insurers and utilities. The agent set those aside, since no company news explained them, and said so. On other days it folds several stocks riding one story (five airlines on an oil-price drop) into one name, so the shortlist doesn’t repeat one idea.

The quality checks weigh how big the earnings beat was and whether the outlook went up, one-time items, the price trend against its 20, 50 and 200-day averages, standard momentum indicators (RSI and MACD), trading volume against normal, analyst target changes, valuation against close peers and the sector’s trend. Ranking puts the strongest company news with the heaviest market confirmation first.

What does the agent actually write?

It writes one record per market day and one record per shortlisted ticker, in the same shape moneyheap’s app already uses for its analysis. Each ticker gets sentiment per time horizon, confidence, reasoning, the case for and against, and what would change the view.

The day’s record holds:

• the date (the New York date) and the trading session the data reflects,
• the standout ticker with a one-sentence reason and a longer one,
• the funnel counts and the method in prose,
• the day’s biggest gainers and decliners among large US companies, each with a sentence or two on why it moved and the sources,
• a short read of the market as a whole: a headline, what is moving it, and the scheduled events ahead,
• the time of the market data snapshot,
• a list of limitations: what the analysis couldn’t know or check,
• a disclaimer and a run status.

Its prompt carries the same written score scale as moneyheap’s other analyses, so a score means the same thing here as on a stock in your portfolio; how moneyheap analyzes the market explains that scale.

Each candidate record holds the price and move at the snapshot time, its role (standout or alternative), why it made the list, and a full analysis: sentiment for the short (about 7 days), middle (1 to 3 months) and long (about 6 months) horizons, each with confidence and reasoning, the news it used, and an assessment with supporting evidence, counter-evidence and an “invalidation” note, the condition that would make the read wrong.

The limitations list is our favorite part. On 2 November it included notes such as: prices are a snapshot one hour into the session, so volume comparisons understate the day; one company reports after the close, so its scores may be outdated within hours; and social data mostly reflects retail forums, so some names lean more on news and analyst tone. The page shows the list under “What this analysis can’t see”, so a reader gets the brief and its blind spots together.

How do we make sure a half-finished run never reaches readers?

The agent marks the day as running when it starts and as completed only after every candidate is written. The email waits for that status, the app shows a day only once its standout is filed, and the email is claimed once per date, so a retry or a slow run can’t send a partial or duplicate email.

• Run status. The page and the dashboard show a day only once its standout is filed, and the archive of past days links only completed days. The email job checks every half hour on weekdays, from 10:00 to 15:30 New York time, and waits until the run is completed and the standout’s own record exists.
• One email per date. Before sending, the email job creates a claim record for the date. Creating it fails if it already exists, so a second run of the job does nothing, even if the agent later rewrites the shortlist.
• Narrow write access. Anyone can read the shortlist, because it is public research. Only one dedicated agent account, marked with a special permission in our authentication system, can write it, and the database rules check that permission on every write. The shortlist goes to every subscriber’s inbox, so it must never be open to public writes.
• Credentials stay out of the prompt. The agent signs in as that restricted account. The hosting platform keeps its credentials and supplies them only to the requests that need them, so the system prompt and the session transcript never contain them.
• Access checks don’t touch real data. When we test that the agent can still write, it writes, reads back and deletes a throwaway record, never a real day’s shortlist.

What would we tell someone building a research agent like this?

Treat the agent as a separate service with a fixed data format, pin its version, make it write down its own limits, and make everything that reads its output safe to run twice.

• Agree on a contract before the prompt. Our app already had a type for an asset’s analysis. Asking the agent to fill exactly that shape meant the page, the dashboard and the email worked on day one.
• Pin the version. A prompt change is a code change. Pinning the deployment to a version gave us the same review step we have for code.
• Ask for counter-evidence and limits explicitly. Without a field for them, a model writes the case for. With a required field, it writes the case against too, and it is often the most useful paragraph.
• Make readers wait for “completed”. Agents take minutes and sometimes fail halfway. A status field and a claim-once email cost little and remove a class of embarrassing bugs.
• Name things neutrally from the start. The feature started out with “pick” in its name, which read like a recommendation. We renamed it the Daily Standout, and the agent writes “why it stood out”, not “why to buy it”. moneyheap is research, and the words should say so.

Frequently asked questions

Does the agent trade or manage money?

No. It reads public market data and news and writes a research brief. It has no access to brokerage accounts, and moneyheap doesn’t place trades.

Is the Daily Standout a recommendation?

No. It is the stock that stood out in that morning’s research, with the evidence for and against it and the limits of the analysis. Momentum describes what already happened and can reverse.

Why run at 10:20 and not at the open?

The first minutes after the open are noisy. Running about an hour in lets the opening burst of trades pass while the move is still fresh, and the agent notes that the volume it sees is still only part of the day.

What happens on market holidays?

The schedule runs on weekdays only. If no shortlist is filed for a date, no email goes out that day, and the Daily Edition page keeps showing the most recent one.

Do I need Pro to read it?

No. The Daily Standout’s full research and the daily email are free. The full analyses of the other shortlisted stocks, the biggest movers and the trending stocks on the same page are part of moneyheap Pro.

Can I ask the agent about my own stocks?

Not this agent; it researches the whole market once a day. For your own watchlist, moneyheap analyzes each asset in your portfolio, and each stock has a public page, for example NVDA. On Pro you can also ask Claude about them over MCP.

How we wrote this

This post describes the Daily Standout research agent as of 6 November 2026 (first written on 2 November), based on our own setup, the data contract and the pages in moneyheap’s codebase, and the shortlists the agent filed on 2, 4 and 5 November. The funnel counts come from those runs and the limitations quoted above from the 2 November run; other days differ, and the brief keeps changing. We left out account names, internal identifiers and credential handling details on purpose. For how moneyheap’s standard analysis is produced, see the methodology, and for the risks of acting on momentum or AI-generated research, read the risk disclosures.

moneyheap is a market research tool. This post is for information and education only. It is not financial advice or a recommendation to buy or sell any security. Do your own research.