The question
The owner says what he wants to know. An agent asks him focused questions, one at a time, sets out the options and writes a design spec with his decisions in it. Nothing moves until he approves it.
Case study · Agentic research
Research with AI agents. A person decides.
We research the stock market with AI agents. The agents write the code, run the studies and check each other’s work. A person sets each question, approves each plan and decides what goes forward. This page shows the method and leaves out every result.
- AI agents
- Write the code, run the studies, review the work
- A person
- Sets the question, approves the plan, decides
- On paper
- Rules-based programs, after every close
- Shown here
- The method only. No results or settings
This page’s own structure, drawn at build from its HTML.
People read the top layer, crawlers follow the middle one and machines parse the bottom one.
20 headings · 52 links · 5 schema nodes
◆ BuildThe loop
From a question to a written verdict.
Studies now follow one written workflow with gates in it. AI agents do the work between the gates. The research’s owner approves each gate before the next step starts, and only he can shorten the workflow.
A test-first plan
Agents design the structure, then write a plan that names the exact files and the tests to write first. The owner approves the plan before any code is written.
Built against known answers
By default, a fresh agent builds each task against tests on synthetic prices (invented price data built for testing), where the right answer is known in advance. A reviewer agent checks each task. A published strategy is rebuilt rule by rule, and each rule cites its source.
Run behind guards
Then the study runs behind guards meant to stop us fooling ourselves. They’re listed further down this page.
An independent review
A reviewer agent checks the whole change, first against the spec and then for quality. The owner reads the findings and chooses which ones to act on.
The verdict, written down
The result goes into a findings log, failures included. A near miss isn’t adopted after the fact.
The owner chooses
From the tested results, the owner chooses what goes onto paper and which changes are adopted. The agents recommend. He decides.
On paper, then Loep
After every close, a scheduled job checks its code against the research, rebuilds the paper accounts and works out the orders for the next open. The results are emailed. A result that holds up can also reach Loep as an indicator. So far, one has.
Who decides
Who does what at every gate.
No AI model picks a strategy, and none runs in the paper programs. The agents write, test and check. The owner chooses at every gate.
| Step | AI agents | The owner |
|---|---|---|
| Question | Ask focused questions and write the spec | Approves the spec |
| Plan | Write a test-first plan | Approves it before any code |
| Review | Check the whole change | Picks the findings to act on |
| Choosing | Recommend from the tested results | Decides what goes forward |
| Each close | Built the programs. No model runs in them | Can hold the day’s run |
Guards
Guards against fooling ourselves.
A backtest can flatter an idea in quiet ways. A company that later failed drops out of the data, or a setting gets picked with the answer in view. Each guard below closes one of those gaps.
- The later studies use the index as it stood on each day, companies that later left it included. A stock can be bought only on days it was really a member.
- The survivorship bias (the skew from studying only the companies that survived) was measured, not assumed, by testing companies that were delisted, failures among them.
- Prices include dividends and are adjusted for splits, so a split never looks like a crash.
- An idea is measured against a blind control (the same test run without the idea) on the same days. The control keeps the same exits and holding rules and enters without the signal.
- Settings are tuned on earlier years only. Each test year runs on settings chosen from the years before it.
- A proposed change is written into the spec before the run, so the test can’t be shaped to fit the answer.
- A result has to hold in both halves of the history and for every random seed.
- Strategy tests and paper records charge realistic trading costs.
- A look-ahead test cuts the price history off at a given day. Nothing decided before that day may change.
- Most ideas fail, and the findings log says so.
- Each column is a trading date and each row is a place in the index.
- Companies join the index and leave it over time.
- The point-in-time record keeps every company for the dates it was a member, the ones that left included.
- A survivors-only view keeps today’s members and drops the rest. The gap it leaves is the bias.
On paper
Followed on paper after every close.
The owner chooses which results go onto paper. From then on, automated, rules-based programs follow them after every market close. AI agents built and reviewed these programs, and no AI model runs in them.
They follow the strategies our research judged most promising. Those are published strategies the agents rebuilt and tested, with corrected versions beside the originals where the published code didn’t do what it said. It’s research, not advice.
Checked before it publishes
The programs run the research’s own code. Where a program uses its own version of a strategy, each run first checks that version against the research. A run also refuses to publish when too little of the market has that day’s prices.
Rebuilt from the start
The paper accounts are rebuilt from their start at every close, and a check reports where the record changed since the last run. The part from before the accounts existed is a replay, worked out after the strategies were chosen.
Orders, never sent
After each close the job works out the orders for the next open. Nothing goes to a broker. The orders are published for a person to read, and the results are emailed.
A person can stop it
The owner can hold any day’s run, and its email with it. He has done so.
Into Loep
Where Loep comes in.
Loep is a research tool we built for people who aren’t professionals. The indicator Loep offers today is a system our research replicated, tested and revised, working with AI agents. The ones still in testing go through the same process.
Nothing carries research into Loep automatically. The indicator’s sheet writes out its method and its weaknesses, and its settings stay out. The paper records don’t appear on Loep.
Withheld from this page
Not shown
What this page leaves out.
This page describes how the research works, not what it found. A single quoted result would be judged without its method, so none appears here. These stay private.
- Every result and figure, and any comparison with the market.
- The names of strategies and stocks, and every setting and threshold.
- What the paper accounts hold, and when their record starts.
- The tools, vendors and published sources the work draws on.
Built under
Next
Start here
Tell us what’s slow, or what can’t be found.
One paragraph is plenty. Tell us what you want built, the step that costs the most, or the page that should rank and doesn’t. If you know why the obvious fix hasn’t worked, say that too.
Email
daniel@adjlcapital.com
Daniel reads and answers these himself, usually within a day or two. No sequence, no drip, and nothing sold to anyone.