A long session with an agent builds up a pile of readings. The levels from when you sat down. A volatility rank from mid-morning. A funding rate from before lunch. All of it stays in the conversation. When you ask a new question, the agent can answer from that pile without making a call. Sometimes that is what you want. With market data it often is not.
Why the agent reuses
A model treats everything in the conversation as available fact. If the SPY call wall is already on the page, a second request for it looks answered. Calling the tool again seems redundant. So the agent repeats the earlier figure, and nothing in its tone tells you it did not look.
The age was true once
Every NoVo reading carries age_seconds. That figure is how old the reading was when the response was built. It does not keep counting inside your chat. An hour later, the agent may quote the reading as a few minutes old, because that is what the payload said. The reading is now an hour older than that. Why the field exists is in every number we publish carries its age.
The honest reference is as_of, the moment the reading describes. An agent reusing a figure should give that time and let you do the subtraction.
Where it bites
It bites when you are comparing. Ask for the levels now against the levels this morning, and a careless agent compares the morning’s reading with itself. It reports no change, and it is right about the two copies of one reading.
It also bites after an event. A release lands, you ask where things stand, and the agent answers from before the release. The answer reads as current. Whether a reading is fresh enough is a judgment only you can make, and you need the time to make it. The general test is in how to tell whether market data is stale.
Make it look again
Say so in the question. Fetch it again. Call the tool now. Those words remove the choice. Add the usual tail: tell me the time of this reading. If the time matches the earlier answer to the second, on data that should have moved, ask whether a tool was called.
For standing use, one rule covers it. For any market figure, make a new call unless I say to use the earlier one. When you do reuse an earlier reading, say that you are reusing it and give its time.
When reuse is right
Reuse is fine for data that has not changed. A weekly positioning report is the same all week. A calendar for the day does not need fetching five times. Reading again there only spends requests. The balance is covered in an agent built around rate limits. The rule is to ask no faster than the data changes, and no slower either.
Start fresh for a new session
Long conversations have a second cost. Instructions given at the top get followed less closely as the page grows. If the agent starts dropping ages or caveats late in a session, open a new conversation. Carry over only what you need, with its times attached.
Every reading on the MCP & API is stamped when it is served. The stamp does not update itself in a transcript. That part is up to the agent, and to you.