Build1 publisher3 min readPublished
Gate 1 halts the AWS design until someone produces peak RPS and an RTO in minutes
The AWS Solutions Architect Prover asks for no AWS credentials, which is the right call for blast radius and also the reason it cannot see the idle NAT gateway its author says models keep missing.
The Engineer · Build desk
What happened
- Renato Marinho argues on dev.to that an LLM asked to build on AWS picks what is popular in its training data rather than what is cost-effective or secure for the constraints in front of it.
- His answer is the AWS Solutions Architect Prover, an MCP server that runs a proposed design through five gates from the Well-Architected Framework in a fixed order and rejects it at the first failure.
- The server asks for no AWS credentials and no access to the environment, because its author holds that handing an agent AdministratorAccess causes the breaches the tool exists to prevent.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Without a credential the tool can only audit the drawing, so the Gate 4 total is a forecast of a new invoice and the resources already accruing charges in your account stay outside its reach.
- decision The adoption cost lands on whoever owns the requirements, because a peak RPS and an RTO have to be committed to before design starts, which moves the argument from the model to the product conversation.
- cost Cross-AZ transfer at a cent a gigabyte only justifies a redesign in the tens of terabytes, so on Gate 4's list it is the fixed items that decide whether an estimate lands anywhere near the bill.
The ordering is the whole mechanism. Gate 1 will not look at a topology until peak RPS, p99 latency in milliseconds, an availability SLO and RTO/RPO in minutes exist as numbers [10]. That inverts the usual flow, in which the model draws first and you argue about the invoice later. What the post describes, though, is a text protocol: the server returns a rejection and demands concrete fixes [9]. No hook into a terraform plan is described, and nothing outside the calling model's own compliance keeps Gate 3 from being skipped [3].
Then the arithmetic. The post's opening waste figure is a $32-a-month idle NAT Gateway that the model does not know is already sitting in your VPC [4]. One per AZ across three AZs is $96 a month, or $1,152 a year, before a packet moves [1]. The other charge the post calls silent, cross-AZ transfer at $0.01/GB [5], is smaller than its reputation at low volume: a terabyte a month costs $10.24 [2]. It reaches $1,024 a month at 100 TB [2]. For either number to transfer you need the same region and the same idle assumption, and the post names no region [3].
The blind spot follows from a design choice I would still make. The Prover requests no AWS credentials and no access to your environment, operating entirely at the reasoning layer [7], and Renato Marinho's stated reason is that agents holding AdministratorAccess produce the breaches the tool exists to prevent [8]. Fair. But a reviewer that cannot authenticate also cannot enumerate, so the total cost assembled at Gate 4 out of compute, storage, NAT gateway, cross-AZ transfer, egress, VPC endpoints, CloudWatch and support tier [13] is a forecast of a greenfield invoice, not a reconciliation of yours. The idle gateway in the opening example is invisible to the Prover for the same reason it is invisible to Cursor: nobody is reading the account [4]. Slightly awkward for a tool that opens with a cost you cannot see from the diagram.
The three named pathologies, requirement hallucination, cost hallucination and service sprawl [6], match what I see: models answering "scalable" with EKS, Kinesis and Aurora Global Databases [2], and proposing multi-region for a tool used by twenty people [16]. That is a prior, not a measurement. The post offers no count of designs reviewed and no rejection rate; the frequency claim rests on the author having watched it happen dozens of times [15][3].
Gate 3's questions are the ones a senior reviewer asks anyway: justify EKS over Fargate, justify Kinesis over SQS [12]. Encoding them buys you the boring Tuesday when nobody asks. That part is worth lifting into your own prompt template whether or not you install the server, and the bill reconciliation stays a separate job for something with read access.
What to watch
- A published rejection rate across real submitted designs, which would turn the author's 'dozens of times' into a measurement.
- Whether a later version takes read-only billing access so Gate 4 can reconcile against an actual invoice rather than forecast a greenfield one.
- Whether the gate order gets enforced anywhere other than the calling model's goodwill, such as a check that runs before terraform apply.