The Subscription Agency, Exposed.

This is an investigation, not a rant. The Subscription Agency (TSA) and the Subbly AI website builder are presented to founders as a fast, low-cost path to a live subscription business. What follows is the documented reality - built from first-hand usage of the Subbly builder, direct quotations from Subbly's own support staff, and the reported experiences of named clients.

Throughout, four categories of statement are kept visibly separate: documented facts and direct quotations (labeled as such); first-hand usage observations; reasonable technical inference drawn from observable behavior; and conclusions about motive, which are never stated as fact - readers are invited to evaluate those themselves.

The Subscription Agency presents itself as a subscription growth partner. In practice, its commercial model is structurally aligned with a single platform: Subbly. TSA builds only on Subbly. It does not evaluate whether Shopify + Recharge, WooCommerce, Kajabi, Circle, or a custom build would better serve a given business - because its revenue depends on moving founders onto Subbly's monthly plans.

This is a documented structural fact, not an inference about intent: TSA's service offering is scoped to Subbly builds. A founder arriving with a business model better served by a different stack will never hear that recommendation. The incentive runs the other direction.

Why the alignment matters

A platform-first agency is paid by the platform's adoption, not by the founder's revenue growth. The two goals coincide only when Subbly happens to be the right answer. When it is not, the founder bears the cost - in monthly platform fees, in a stack that fights their business model, and in growth work that never gets done because the engagement ended at launch.

Subbly once operated a certified Experts Program - a network of vetted partners who built and maintained subscription sites on the platform. The program gave founders a marketplace of independent operators to choose from, and gave those operators a credentialed channel for work.

That program was shut down. What follows is the documented sequence; the conclusion about motive is left to the reader.

• 1TSA and Subbly built a certified partner / expert network.
• 2The program was quietly discontinued.
• 3Clients who had been served by those partners were retained inside TSA's own service funnel.
• 4That retention generated a new, direct revenue stream from ongoing build and maintenance work.
• 5The same funnel continued to feed the Subbly platform adoption pipeline - every retained client stayed on Subbly's monthly plan.

This section is self-contained. It is referenced again, briefly, in the conclusion.

Four case studies follow. Each is presented as a documented client experience with the specific cost or commercial anomaly that defined it. Where a figure is a first-hand observation, it is labeled as such.

Case 1 - Cozy Crafter Crates (Charlotte): the $5 vs $0.42 image

Charlotte, founder of Cozy Crafter Crates, used the Subbly AI builder to generate her subscription box site. During the build she prompted the builder to produce imagery for her product pages. The builder's agent loop consumed credits generating and regenerating images. The observed cost for a single usable image generation cycle was approximately $5 in consumed builder credits.

Case 2 - The $11.75 vs ~$2 comparable-work comparison

On a comparable task - a section of work a competent operator or a direct model call would complete for roughly $2 - the Subbly builder's credit consumption was observed at $11.75. The delta is attributable to repeated context loading, rereads, and agent re-entry into the same files.

Case 3 - Pure Roasts: the $26.90 'Fix with AI' incident

Pure Roasts, a coffee subscription founder, encountered an error in the builder. The product surfaced a 'Fix with AI' button. This was product-initiated - the button was offered by the builder UI, not requested by the user. The founder clicked it. The resulting agent loop consumed $26.90 in credits attempting to resolve the error, and did not resolve it.

Case 4 - The coaching client billed $5,000 after being told she had 'won'

A coaching founder was told she had 'won' a free website build through a TSA promotion. She was subsequently billed $5,000. The discrepancy between the representation ('won', 'free') and the commercial outcome ($5,000 charged) is the documented fact. Whether it constitutes a misrepresentation is for the reader to judge.

The case-study figures are not isolated. They are the visible surface of a structural cost pattern in the builder's agent architecture. A direct model call charges for exactly the tokens used. An agentic loop charges for every step: loading context, reading files, reasoning, calling tools, re-reading the same files, retrying, and re-entering context after compaction.

When the loop is well-designed, the overhead is justified by autonomy. When it is not - when the agent re-reads files it has already read, retries without new information, and re-loads context it just compacted - the founder pays for the builder's inefficiency, denominated in credits.

The observed multipliers - 5.9x and 12x - are consistent with the first pattern, not the second. They are first-hand usage observations, not estimates.

'Agent thrash' is the term for an agent that re-enters a loop without making forward progress: re-reading, re-reasoning, re-calling the same tool, consuming credits without converging on a result. The Pure Roasts incident is a textbook case.

The product offered a fix. The fix ran. It did not fix. It cost $26.90. The founder, who did not initiate the action, absorbed the cost. This is the core of the problem: a product-initiated action that bills the user for its own failure to resolve.

The Subbly AI builder generates a codebase. Subbly's own support staff have stated, verbatim, that maintaining that codebase is the user's responsibility.

This creates a contradiction the founders are not warned about. The builder is sold as a no-code path to a subscription site. The output, however, is a codebase - and the responsibility for that codebase, including its bugs, its drift, and its technical debt, is transferred to a founder who chose a no-code tool precisely to avoid maintaining code.

The technical debt is not hypothetical. Each generation produces residual code - dead branches, unused imports, half-implemented logic the agent abandoned. That code accumulates. And per Subbly's own statement, the founder owns it.

When confronted with these observations, Subbly's responses are worth reading verbatim. They are reproduced here without paraphrase.

Read that carefully. The stated position is not 'we will fix it.' The stated position is that the issues will persist as long as the founder works with the consultant (Alex) - framing the problem as a function of the consultant's involvement rather than the platform's behavior. Readers can evaluate whether a platform that conditions its own correctness on the absence of a specific external party has demonstrated the platform works.

The same support thread that asserted 'nothing changed' also acknowledged the deployment of advanced chat compaction - a material change to how the agent manages context. The two statements are in direct tension. 'Nothing changed' and 'we deployed compaction' cannot both be literally true of the same system in the same window. This is documented, not inferred.

When the cost anomalies and thrash were raised, Subbly's Andrija offered an analogy to explain the unpredictability.

Note the move. The comparison the founder drew - same kind of work, two different days, a large cost swing - is reframed as invalid because the runs were 'not the same page, not the same referenced context window, not the same referenced codebase.' That is precisely the point. A builder that cannot hold context across two comparable tasks, and so charges multiples more for the second, is the platform problem under discussion. Reciting the conditions that produced the cost swing is not a refutation of it. It is a description of it.

The analogy is technically true and analytically useless. LLMs are nondeterministic at the token level - that is a fact. But the basketball analogy is deployed to explain away a 5.9x and 12x cost variance, and a $26.90 fix that fixed nothing. Token-level nondeterminism does not produce order-of-magnitude cost swings. It produces small variances around a stable mean.

A 12x multiplier is not 'a slightly different bounce.' It is a structural difference in how many steps the agent took - how many files it re-read, how many times it reloaded context, how many retries it issued. Those are engineering decisions in the agent loop, not statistical noise from the model.

Advanced chat compaction is a technique for managing a long agent conversation by summarizing and compressing earlier context so the conversation can continue past a model's context window. It is, legitimately, a real engineering response to context limits.

But compaction has a cost the founder pays for. When context is compacted and the agent later needs a detail that was summarized away, the agent re-reads the source files to recover it. Each reread is a tool call. Each tool call consumes credits. A builder that compacts aggressively and then rereads to recover lost detail will cost more - not less - than one that managed context cleanly from the start.

This is the technical inference that connects 'advanced chat compaction' to the observed cost multipliers. It is not proven by the quotes alone - it is a reasonable inference from how compaction works, consistent with the 5.9x and 12x figures. It is also why the 'nothing changed' assertion is so difficult to sustain: deploying compaction is, by definition, a change to how context is managed, and the downstream effect of that change is exactly the kind of reread-driven cost the observations show.

The debate over prompt quality versus platform problem is, at present, an argument about rhetoric. It should be an argument about logs. A builder that instruments its agent runs can answer every disputed claim with data. The questions a serious builder's telemetry should answer for any given prompt are specific and finite.

• 1Context loaded: how many tokens of context were loaded, and from which sources, before the first reasoning step.
• 2Reasoning steps: how many discrete reasoning iterations the agent executed before terminating.
• 3Tool calls: every tool invoked, in order, with the arguments passed.
• 4Files accessed: which files were read or written, and how many times each was read across the run.
• 5Cache hits: which reads were served from cache versus re-fetched from source.
• 6Fallback models: whether the run fell back to a different model, and why.
• 7Retries: how many times a failed step was retried, and whether the retry used changed or identical inputs.
• 8Reread behavior: whether the agent re-read files it had already read in the same run, and after compaction.
• 9Termination cause: whether the run ended on success, a stop condition, or credit-budget exhaustion.

If the telemetry shows the disputed prompts triggered excessive rereads, retries with identical inputs, and termination on credit exhaustion rather than success, the platform problem is demonstrated in data - not analogy. If it shows clean single-pass runs with stable costs, the prompt-quality defense is vindicated. Either way, this is where the conversation belongs. A builder that will not expose these logs is asking founders to take its word over their invoices.

Strip the diplomacy from Subbly's support thread and the sequence is blunt.

Read the moves in order. The founder reports a problem with the builder. Support attributes the problem to the consultant, not the platform. Support states the problem will not be fixed as long as the consultant is involved. Support then offers Subbly's own team to fill the role the consultant was performing. 'Not a pitch' is, structurally, a pitch - the platform positioning itself as the replacement for the partner it just blamed.

Step back and the shape is the same in both cases. The Experts Program (Section 3) eliminated Subbly's independent partner layer and absorbed those clients into its own service arm. The managed-arrangement offer in Section 12 tries to do the same thing to this founder - replace the outside consultant with Subbly's internal team.

Two moves, one direction: the service layer that once sat between the founder and the platform is being pulled in-house. The platform that builds the tool, bills for the tool, and answers support tickets for the tool does not need to also run your build to be useful - but it is structurally cheaper for it if it does. A founder building on a single owned stack should know that the stack's owner is also bidding to be the builder.

There is one more move in the support thread, and it is the ugliest. The founder asked why her costs swung. Support did not answer the question. Support reassigned it.

The founder is told, in writing, that the question about her own bill is 'not coming from you' - that it is really the consultant's question, not hers. The function is obvious: separate the founder from the person raising the platform's behavior, and the behavior can be reframed as someone else's agenda.

Except the founder noticed first. She did the work herself, watched the cost herself, and wrote to support herself - before anyone framed it as anyone else's question.

When reassigning the question did not make it go away, support switched tactics - from 'that's not your question' to 'you don't understand your own question.'

There is one more piece of evidence, and it is the simplest to verify. The Subscription Agency's site publishes a 'Client Stories' wall of testimonials - eight cards, each with a five-star review, a name, a role, and a smiling headshot. As of the date of writing, the names and the faces do not hold up.

Documented · The Subscription Agency - Client Stories section

The headshots are the first tell. They are the generic, symmetrical, evenly-lit portraits that AI website builders emit by default - the same handful of synthetic faces that recur across AI-built sites. They are not photographs of real clients. They are default avatars dressed up as social proof.

The names are the second. One card credits 'Lisa Gotravel, Founder, WanderKit.' WanderKit is a real company - but a search for its founder returns a different person entirely.

Documented · Google - Lisa Gotravel Founder, WanderKit

The name on the testimonial does not match the real founder of the real company. 'Lisa Gotravel' appears to be a pseudonym - a handle, not a person - attached to an AI-generated face and presented as a verifiable client. The other names on the wall follow the same pattern: plausible, generic, and unattached to any company you can call.

The case studies, the cost analysis, the verbatim quotes, and the commercial sequence are now on the record. The thesis that closes this investigation is stated once, in full.

This is not a claim that Subbly is useless, or that TSA acts in bad faith. It is a claim that the defenses offered - nondeterminism, prompt quality, the consultant's presence - do not account for the documented costs, the verbatim contradictions, or the structural incentives. A founder deciding where to build their subscription business deserves that record before they commit.

If that founder wants a genuinely platform-agnostic, operator-first partner - one that recommends the right stack for the business, engineers the retention work that actually compounds, and stays past launch - that alternative exists. It is what Subscription Scale is built to be.

Frequently Asked Questions