Pricing

The scan is free because it costs us nothing. The audit is not, because it cannot be automated.

Everything a scanner can check, you can have for nothing, including the formula. What you pay for is the part that requires running real agents against real code and reading what they did.

Free scan

instant, no account

What a machine can see from outside.

15 deterministic checks across five funnel stages, 17 points on paper
A score out of the points we could actually measure on your domain, with the rest named rather than counted against you
A permanent link you can forward
The published formula, so the number can be argued with

Scan your domain

Diagnostic

$2,900

one week

One behavioural cell, not a better version of the free file.

Three real agent runs on one brief we design with you, recorded
Whether you are in the candidate set at all, or absent from it
The words used to reject you, quoted where the run left a quotable sentence and labelled as our summary where it did not
Sixty minutes to walk through it with whoever owns the fix

Credited in full against a Full audit booked within 90 days.

Ask about a diagnostic

Full audit

$11,000

two to three weeks

What agents actually do when nobody is watching.

Three experiment cells (model × condition × brief), four runs each, recorded
A brief designed for your category: on the same cheaper model, one brief produced 0 of 10 runs that read any documentation and another produced 3 of 3, because the second decision turned on a licence
Which provider gets picked over you, and the words used to reject you
Where an agent stalls: registration, credentials, or the first integration
A re-measure after 60 days on the same brief, the same scaffold and pinned model versions, with any model change reported as a confound rather than as a result

Ask about an audit

Why one run of an agent proves nothing

Ask a model the same question twice and it will not answer the same way. Our own runs say it plainly: one vendor was rejected in nineteen of twenty runs and chosen in the twentieth, and one signup endpoint answered 200 once and 403 four times inside an hour. Anyone selling you a position in an AI ranking from a single run is selling noise, and they should be treated that way.

So we do not sell a position. Every cell is repeated, every run is recorded separately, and what you get is the spread rather than a number: how many runs chose you, how many named you at all, and the words each one used. Where runs disagree, the disagreement is the finding and it is printed as one. The four published audits are written that way, and you can check that before paying us anything: each one names the models, the number of runs, the scaffold, and what a different scaffold would have changed.

The honest limit is the sample size. Four to six runs per cell is enough to see a wall every run hits and not enough to rank two vendors that finish close together. We report which of those two a finding is, every time.

Read the four published audits, which are the sample of the deliverable

Who does the work

One person, and you should know that before you weigh a five-figure engagement. Krystian Gwizdała, a full-stack engineer in Kraków, wrote the scanner, ran all eighteen agent runs behind the published audits, and will be the one reading your transcripts. There is no team behind this and no account manager between you and the work.

What that costs you is capacity: two full audits a month, not ten, and no cover if one week goes badly. What it buys you is that the person who designed the brief is the person who argues with you about what the runs mean, which is the part of this that does not survive being handed over.

After the findings

An audit that ends in a document changes nothing. These are scoped from what the audit found, so the work is already argued for before it starts.

Fix sprint

$7,500 to $16,000

Implementation, not advice: documentation that answers the question agents actually ask, llms.txt worth reading, an entry point built for a machine, a credential path that does not need a human.

MCP build

$14,000 to $28,000

An MCP server for your API, scoped from the audit rather than from a wishlist. MVP or a standard build with OAuth and write access.

Retainer

$3,000 / month

Quarterly re-measurement, alerts when your score moves, and priority access. Only worth buying after an audit, because before one there is nothing to compare against.

Who runs this

Krystian Gwizdała. The research on this site is mine: the scanner, the formula, the agent runs and every number on Findings and the industry report. An audit is run by me, not by a team you never meet, and the raw runs go to you with the report.

hello@letagentsin.com

Questions people ask first

Do you bill hourly?: For work outside a package, $250 an hour with an eight hour minimum. It is rarely the right shape: the value here is a measurement and a decision, not time at a desk.
What if the audit finds nothing?: Then you get that in writing, with the runs to back it up, and you stop worrying about it. A report that cannot come back empty is not a measurement, so that outcome has to stay on the table.
Why is the free scan actually free?: It is HTTP requests. No language model runs, so it costs bandwidth and nothing else. Charging for it would be charging for a script whose every rule is published, which you can reproduce with curl.
Will you publish what you find about us?: The scans in our own published corpus are published as we produce them, because every line is one HTTP request with a published rule and you can reproduce all of it. A scan you run yourself never joins that corpus: it gets a permanent link you can forward, and we do not post it anywhere. A paid audit is yours: nothing from it is published without your written agreement, and if we ever want to write about a pattern we saw, you get the draft and ten working days before anything goes out.
What do you need from us?: For the audit, nothing but the product as a customer sees it. No repository access, no staging environment, no calls with your engineers. That is the point: the measurement has to happen from outside, the way an agent meets you.
Can we just buy the fixes?: Yes, but a fix sprint without a measurement is guesswork with an invoice attached. If you already know what is wrong, say so and we will scope it directly.