A PM spent thirty minutes debugging a table ChatGPT invented for him

A model with no connection to the database cannot check a table name against anything, so it returns one that fits the shape of the query. The author wrote that the fault is not the model's; it sits in the missing harness ^[3].

For that case he prescribes a server. A product harness exposes authorized views through an MCP server, with an allowlist of queries, PII masking and limits on time and rows, so a PM gets business tools instead of arbitrary SQL ^[15]. He notes the PM did not know a database MCP existed ^[16]. Standing that up is an access-control review with a protocol in front of it, and it needs whoever owns the schema.

The token figure is one run. Directed context took an investigation from about 50k tokens to about 15k with the model and the task held constant ^[11], which is 35k fewer tokens, or roughly 70 percent ^[20]. He says he has not treated it as a benchmark, calling it an observation rather than a systematic measurement ^[12]. Each steering loaded is one question the agent does not ask again, he writes, because an agent that already knows the backend is PHP 8 stops asking ^[21]. For the ratio to hold on someone else's workload you would need a schema big enough that pasting it dominates the prompt, a query pattern repeated often enough to be worth wrapping in a tool, and a server whose responses are genuinely bounded. Tool design, restricted schemas and response limits decide that last one ^[13].

In my view the permission split is the first thing to copy, because it outlives whichever model you are using this quarter. The example in the post is specific: the agent may run npm test and may not run rm -rf, may read the database and may not write to it ^[7]. On the difference between a basic and a mature harness he is blunt: in the basic one you trust, in the mature one you audit ^[10].

Before the shared harness, his team's four frontend repositories were each configured differently, and pull requests went back and forth over the inconsistency ^[17]. Shared configuration is a reasonable fix for a review-queue problem.

The fifth section of the post is a checklist to run before granting autonomy, and the text breaks off before the items ^[19]. Model routing gets one line of policy: simple tasks to the cheap model, complex tasks to the expensive one, decided by the harness and not at each prompt ^[14].