Agent Integrity and Efficiency
Your agents do not fail loudly. They fail expensively — burning tokens deciding what to do, fetching the wrong thing confidently, and committing to plans built on something that was never true. None of it raises an exception. All of it is on your bill.
Two things. It measures the spend that produced nothing usable — a waste rate with the denominator shown — and it prevents that spend, by selling the cheap outside check an agent should make before an expensive commitment. Checks are priced from $0.20 and paid per call over HTTP 402 in USDC: no account, no card, no signup.
The five ways an agent burns money without erroring
| Kind | What it looks like |
|---|---|
| False premise | Competent work, aimed at a world that does not exist. The plan runs to completion and lands on nothing. |
| Wrong target | Right effort, wrong object — wrong file, wrong page, wrong record, wrong endpoint. |
| Overreach | A conclusion the evidence gathered could not carry. Every downstream step inherits it. |
| Rework | Redoing work already done, or undoing its own correct output. |
| Dead call | A call that returns nothing usable — timeout, empty result, unreadable format. |
Your monitoring sees none of these, because in every one of them the run succeeded.
Whysaid keeps the ledger: POST /v1/integrity/event to record one, or the free
report_waste tool over MCP. GET /v1/integrity returns your waste rate
alongside total spend — a rate without its denominator is a number that hides which problem you
have. Waste your agent reports about itself is stored and reported separately from waste we
detect independently, because an agent grading its own work is the thing this company exists to
argue against.
What it looks like measured
Our own agent, all traffic, since the ledger opened. Nine events across 234 metered calls. This is a real measurement of a small sample and it is ours — it is not an industry rate, and nobody should plan against it as one. Your number comes from your ledger or it does not exist.
Spend measured
$8.40
234 metered calls
Produced nothing
$0.65
9 events
Waste rate
7.79%
of measured spend
Cause known
2 of 9
7 not yet attributed
One kind dominates, and it is the one nobody instruments: a false premise. The work was competent every time. It was aimed at a world that did not exist, and the run reported success. That is why it is worth $0.35 to check the premise before the plan rather than after it.
The arithmetic of checking first
A false premise caught before the plan costs $0.35. Caught at the end of the plan, it costs the plan. That ratio is the entire product: an agent cannot verify its own recall, and it cannot tell its own earlier output from a fact it learned, so it needs a second party that goes and looks.
| Tool | What it answers | Price |
|---|---|---|
| verify_answer | Open every source a whole answer cited and return a public receipt. | $1.00 |
| answer_share | Is this brand named in live AI answers? | $0.75 |
| verify_claim | Is the premise true — checked, not asserted — before you commit to the plan that depends on it? | $0.35 |
| check_provenance | Was this text model-generated and seen before — before it enters memory as a fact? | $0.25 |
| verify_source | Is this URL live, and does it actually say that — before you spend a context window on it? | $0.20 |
| report_waste | Record work you paid for and threw away; get your waste rate back. | Free |
How to connect
Whysaid speaks MCP, so the tools appear in any MCP-capable agent runtime, and x402 over HTTP for direct calls. An unpaid request returns a 402 carrying the price, asset, network and destination; retry with a payment authorisation and the work runs. Verification happens before the work and settlement after it, so a caller is never charged for a request that failed.
No custody
Payment settles directly from the caller to the receiving address. Whysaid never holds funds.