给 Jev 喂什么状态：退款决策的 8 步状态包

morph @morpphhhaw原推发布于 2026.09.24

在 X 上打开原推

TypeSafe founder Diogo Almeida introduced Jev, and the field I would inspect first in any agent built around it is not model. It is state An agent can have room for a long conversation and still miss the one line that matters. A refund request needs the current order, the charge record and the policy that applies today. It does not need twelve earlier attempts to sound helpful Here is the state packet I would put in front of Jev: step 1 → name the decision the code needs to make now, not the entire job the agent was given step 2 → pull facts from the system of record: order status, amounts, timestamps and permissions step 3 → include the customer's actual words as evidence, without asking Jev to reconstruct the whole conversation step 4 → state the constraints that can change the route, such as a refund window or an account hold step 5 → build the available actions from live code, so a closed account never appears as a valid destination step 6 → put the question in the question field. State is evidence, not a second prompt hiding instructions step 7 → leave the sums and date arithmetic to code, then send Jev the result it needs to judge step 8 → keep the exact packet with the answer, because you cannot debug a decision from its label alone The point is not to compress everything until it looks clever. The point is to make it obvious which piece of evidence could change the answer The document below shows that packet on one page. The article goes further into shaping state without turning it back into a transcript

这条是引用着另一条写的：x.com/morpphhhaw/status/2102740313953439950