CodeLabProsCodeLabPros — homeMenu

Autonomous systems

An agent that researches sales prospects on its own

An agent that sources businesses, measures what is wrong with their websites, and decides which are worth approaching. Unattended, but not unsupervised.

Our role
Designed and built it
Built with
Python · Mapping APIs · Google Lighthouse · Durable file state
Where it got to
Internal, running
01

The idea

Stop finding clients by hand. Have software do the sourcing, the research and the qualifying, and hand over a short list worth acting on.

02

What it actually needed

Volume turns out to be the easy part. Anything can produce two thousand businesses. What makes such a list useless is that everything survives it, so the filtering you were trying to avoid simply moves to the end. The product is not the finding. It is the rejecting, and the ability to say why each rejection happened.

Runs unattendedSOURCE · MEASURE · REJECTSTOPSpending moneyContacting a real personTouching anything liveA human decides
All three limits were written while the agent could not yet do any of them. Once a capability exists, every argument for restraining it starts to sound like an argument against shipping.
03

What got built

Sourcing that resumes
Works through regions and industries continuously, keeping durable state between runs, so it picks up where it stopped instead of starting again.
Measurement the subject can repeat
Assesses each business's website using public tooling, not its own timing, so any number it quotes can be reproduced by that business in about thirty seconds.
Scoring built to reject
Every business is kept or dropped with a written reason attached, not a bare number, so when a reason turns out to be wrong, the rule behind it can be found and changed.
Three things it will not do alone
It stops and waits for a person before spending money, before contacting a real human being, and before touching anything live.
04

The decisions that shaped it

Write the limits before the capability

The three stopping conditions were decided while the agent could not yet do any of them. Once a capability exists, every argument for restraining it starts to sound like an argument against shipping, and the limits get written after the first incident instead of before it.

Use somebody else's instrument

Its own timing measurement was faster, cheaper and entirely under our control. It was also unverifiable by the person being told about it. Moving to a public tool made the numbers slower to collect and impossible to dismiss, which turned the claim from trust us into go and check.

05

In production

Internal software, running unattended. It rejects most of what it sources, and every business it keeps or drops carries the written reason why.

06

What this is worth

Software that runs unattended and still cannot spend money, contact a real person or touch a live site without asking. The limits were written down before the capability existed.

In more detailfor whoever wants it

This runs for hours without us. It finds local businesses, pulls their sites, measures them, scores them against written criteria, keeps durable state between runs, and gets sharper by reading its own learning log.

Every prospect it keeps carries a written reason and, before any conversation happens, the condition that would prove the whole premise wrong. Rebuilt recently. A vendor owns the site under contract. The phone drives everything and the site is decorative. Cannot spend the money.

It stopped trusting its own measurements

The output is literally a number: your site takes N seconds to load. We were measuring that ourselves: request the page, time the response. Fast, free, and wrong: it measures how quickly bytes leave a server, not how long a person waits for something to appear. On a page carrying sixty scripts those are nowhere near each other.

It uses Google's own measurement now. Slower, rate-limited, more work to wire up. Every number it quotes belongs to Google, so the business can reproduce it in thirty seconds by pasting their own address into a public tool. That turns the claim from trust us into go and check.

It also refuses to put a currency figure on a loss it has not observed. Writing this is costing you $40,000 a year would be easy and it is what everyone does. But any owner who knows their own numbers spots an invented one instantly, and then doubts everything else in the message.

Three things it never does alone

It will not spend money. It will not contact a person. It will not touch a live site. Up to those lines it is autonomous; at them it prepares the action in full, presents it, and waits.

The rule that makes it safeOne approval never carries to the next. Saying yes to one message does not authorise the following one. That is deliberately tedious, because software like this does not fail through malice. It fails by reasoning its way from you said yes once to you would probably say yes now, at three in the morning, with nobody watching. And where it is unsure whether something crosses a line, it treats it as though it does.
Where this leadseach link says why
The principle behind itLet unknown be an answerA system that cannot say 'we do not know' will say something else instead, and you will not be able to tell which it did.What it taughtWhy we stopped using our own performance numbersOur own measurement was faster, cheaper and entirely under our control. It was also unverifiable by the person we were showing it to, which made it worthless for the only job it had.What it taughtWe fixed the same bug twice, two commits apartOne commit says measure the problem instead of guessing at it. The one after says measure with somebody else's instrument, because our own timing was wrong. Same error, one layer down.What it taughtThe three things the agent is not allowed to do without usSpending money, contacting a real person, touching anything live. All three were written down while the software could not yet do any of them.Read nextInfrastructure for research that keeps itself honestFrom an agent that decides what to pursue, to the substrate that decides what may be believed in the first place.If you have one of theseEmbedded Engineer — from $7,500Software that runs unattended is where a written decision record earns its keep, and that is what a retainer buys.