Returning the count from the engine fixed a question a small model missed twenty times out of twenty

The terminal table and the agent tools read the same report, and can print different numbers. According to the article, when they do, one figure is wrong in a way that survives review, because each looks sourced ^[23]. An end-to-end test holds the terminal table and the agent tools to the same number ^[1].

The server states the rule before any tool call. Its `instructions` field, served to the client on connect, reads: "zombiescan finds AWS resources nobody is using and prices them. Every tool is read-only; nothing here deletes anything. Run scan_account once, then ask estimate_savings for any total, count or breakdown of the report it writes -- do not add up findings yourself." ^[10] The instructions field is advisory; what binds is what the tools return: the costliest row is picked in code, the per-check and per-region sections are grouped in code, and the model quotes what came back ^[15].

The result behind the rule is a counting failure. One small model, asked how many of eleven ids were 10 or more, answered wrong in twenty calls out of twenty ^[6], and pushing the count into the engine makes the same question right every time, the author writes ^[7]. The model is not named. For 20-out-of-20 to say anything about your agent, three things would have to hold: the model is comparably small, the question makes it evaluate a predicate row by row, and the rows are in context for it to try. A model handed rows and asked how much they cost does the addition in the answer ^[24].

The totals are the least interesting output here. Thirty-one of the 120 findings carry a price ^[4], which leaves 89 at zero and matches the `free_count` field ^[20]. `estimate_savings` also reports `approximate_count` as 100 ^[14], so 100 of the 120 prices are estimates, and the single costliest finding is an ECR repository in us-east-1 at $0.52 a month ^[14]. One check, `ecr-stale-images`, accounts for $6.40 of the $6.90 monthly figure, about 93 percent ^[19]. The engine returns that qualifier as a field. The demo account's whole waste is $82.80 a year ^[3]. The size of the bill is beside the point.

Adoption cost is specific. There is no MCP SDK in the dependency list, the framing runs to about a hundred lines, and the stated reason is that a second package to keep current is a second thing to break ^[12]. You own the handshake in exchange. The article's advice is to answer `initialize` with the client's own `protocolVersion` when you recognise it, because a client that gets its version back proceeds and one that gets a different string negotiates again ^[13]. The permissions are the other line item: read-only access wide enough to describe resources in every region you scan, plus `ce:GetCostAndUsage` for the calibration step ^[18].