7 questions to ask before your agent acts on a number

Your agent buys data thousands of times an hour and checks almost none of it. It asks an API for a price, a balance, a rate. The response comes back in milliseconds, looking exactly like every other response it has ever seen, and the agent acts on it. A trade fires. A payment settles. A position moves.

A human in that seat would hesitate. They would glance at the number, sense that something was off, open a second tab to confirm. An agent does none of that. It takes the number at face value and acts at machine speed, and if the number was wrong, the mistake is already done before anyone knows a mistake was possible.

The gap is not the data itself. Data is everywhere, cheap and abundant. The gap is that nothing sits between the data and the decision asking whether the number deserves to be trusted. These are the seven questions that belong in that gap, and what it costs when nobody asks them.

1. How old is this number?

A price from five seconds ago and a price from five minutes ago look identical to your agent. Same shape, same field, same confident value. But in a market that moves, the age of a number is the difference between a good decision and a loss.

Staleness is the quiet failure. A feed that updates on a heartbeat, a cached response served after an outage, a signal that aged while it traveled through your pipeline. None of it announces itself. The number arrives looking fresh, and the only way to know otherwise is to carry the moment it was actually measured, at the source, and compare it to now.

An agent acting on a five minute old funding rate is not making a slightly worse decision. It is making a decision about a market that no longer exists.

2. Where did it actually come from?

One source or several. A named exchange or an aggregator. A direct observation or a number that passed through three hands before it reached you. Provenance is the first thing a careful analyst checks and the first thing most data pipelines discard.

The problem compounds with aggregators. A single feed that quietly pulls from one upstream can present itself as authoritative while being nothing more than a relay. When the upstream is wrong, the relay is wrong, and there is no second opinion anywhere in the chain, only the appearance of one.

If you cannot name where a number came from, you cannot reason about how it might be wrong.

3. Do independent sources agree?

Agreement between sources is only meaningful if the sources are actually independent. This is the trap that catches most systems. Five providers returning the same number feels like strong confirmation, until you learn that all five read from the same upstream feed. That is not five observations. It is one observation counted five times, wearing the costume of consensus.

Real independence has to be measured, not assumed. Two exchanges reporting the same spot price from their own order books are genuinely confirming each other. Two services reading the same oracle are confirming nothing. The difference is invisible in the numbers themselves and decisive for how much the agreement is worth.

When independence is measured rather than presumed, agreement becomes evidence. When it is presumed, agreement is theater.

4. Was it measured, or calculated?

There is a difference between a number that was observed and a number that was computed, and that difference disappears the moment both are formatted the same way in a response.

A last traded price is an observation. Something happened, at a size, at a moment. A mark price or an index is a construction, often blended from several venues, sometimes incorporating the very sources you are trying to check it against. Both are legitimate. Both are useful. But an agent that treats a calculated number as a raw measurement is trusting a model it never saw and cannot inspect.

The honest response carries the distinction. Measured means lifted directly from what a source reported. Derived means computed from other values. Passing the second off as the first is a small fiction at the exact moment you most need the truth.

5. How much should I trust it?

The useful question is never whether a number exists. Data is almost always there. The question is how much confidence the evidence actually supports, and that is a question most systems refuse to answer because answering it honestly means sometimes admitting the answer is not much.

An agent needs this as a value it can read and branch on, not as a vibe. Trust it enough to act. Trust it enough to flag for review. Or do not trust it enough to move at all. A confidence level that the agent reads in code turns a raw number into a decision it can defend, and turns a shaky number into a reason to wait.

Confidence that reflects evidence is worth something. Confidence that reflects nothing but the fluency of the response is worth exactly nothing, and looks identical.

6. What happens when the data is bad?

Every data source fails eventually. The question that matters is what your pipeline does at that moment. The dangerous answer is that it returns something anyway. A stale quote. A zero where a real value should be. A best guess dressed as a fact.

A zero is the sharpest example. To an agent, a balance of zero is a real, actionable answer. An unreachable source reporting zero would tell an agent that a funded treasury is empty, right before it decides whether to move money. The failure is not that the source went down. The failure is that the system hid the outage behind a number that looked fine.

The correct behavior when nothing can be verified is to refuse, clearly, rather than to serve a confident wrong answer. A refusal costs a moment. A confident wrong answer costs whatever the agent did next.

7. Can I audit the decision later?

When a trade goes wrong, the first question is always the same. What number did we act on, where did it come from, and why did we believe it. If the answer is a black box, the loss teaches you nothing and you are exposed to the same failure tomorrow.

An auditable decision carries its own record. Which sources were considered, what each one returned, which one was chosen and why, what checks it passed, how fresh it was. Not as a favor to compliance, but because a decision you cannot reconstruct is a decision you cannot learn from or defend.

The number is the smallest part of the answer. The reasoning behind it is what lets you trust the next one.

The gap between the data and the decision

None of these questions are exotic. They are what a careful person asks instinctively before acting on information that matters. The problem is that agents do not ask them, and the systems feeding those agents were built to deliver numbers, not to judge them.

That gap, between the data arriving and the agent acting, is where trust either exists or does not. Today, in most systems running right now, it is empty. The number arrives and the agent believes it, every time.

Metera exists to fill that gap. Every answer it returns carries the age of the data, the source it came from, whether independent sources agreed and whether that independence was measured, whether the value was observed or computed, an explicit level of how much to trust it, and a full trace of how the decision was made. When nothing meets the bar, it refuses instead of guessing, and the refusal costs nothing.

Your agent will act on the next number it receives. The only question is whether anything checked it first.

Metera is the trust layer for AI agents. It picks the best source for a piece of data, verifies it, and returns it with an explicit level of how much to trust it. Live on Solana. One line to connect.

claude mcp add --transport http metera https://api.metera.xyz/mcp