State of AI Agent Verification 2026: What 93 Verifiable Receipts and 30,710 Tool Calls Show

The 93-receipt corpus

The sample is 93 receipts in two files: 53 banked ledger receipts from September 16 to 17, and 40 freshly minted receipts from September 21 to 25. Together they cover a 9-day window and 0.303% of all-time tool calls, so treat this as a sample with full receipts, not a census.

Each receipt carries nine fields: the task, the tool, the input arguments, a result summary, a verifiable hash of the result (result_sha256), a public receipt URL (zambo.dev/run/<id>), a timestamp, a tier, and a source. That schema is identical across all 93 rows. There is no claim-vs-execution field in the schema and no chain-anchoring field in the schema, so this page reports neither: the data does not support those metrics, and they are not invented here.

Daily receipt counts

Weekly: ISO week 38: 41, ISO week 39: 52.

verifiable receipt

A public record of one AI tool call that includes the input, the result summary, a cryptographic hash of the result, and a timestamp, served at a permanent URL. The hash lets anyone check that the published result matches what the tool returned.

result hash

The sha256 digest of the tool's result, recorded on the receipt as result_sha256. In this corpus every receipt carries one: 93 of 93.

The 10 tools that show up in the receipts

Ranked by receipt count in the 93-receipt corpus:

1. live_price: 20
2. provibe_audit: 17
3. ghost_audit_site: 16
4. capability_search: 11
5. zambo_universal: 9
6. prompt_shield: 6
7. leadsignal: 6
8. credithunt: 6
9. ghost_audit_status: 1
10. ghost_audit_report: 1

Price lookups and site audits dominate the sample; capability_search shows up 11 times here. In the live platform totals, the all-time leader is capability_search with 3,819 of 30,710 calls (12.4% of all-time calls).

Across the whole platform, the shape is different from the sample. 15,445 calls (50.3% of all-time calls) are classified as work tool calls, meaning calls doing real work rather than discovery or testing.

work tool call

A tool call classified as doing real work for the user, as opposed to discovery, capability search, or test traffic.

Reliability harness results

The platform runs a reliability harness that records test results per tool. As of the last run on 2026-09-25 08:10:05 UTC, it held 5,797 recorded results across the full 132-tool surface, with coverage marked "full". That means every tool in the catalog has recorded results, not just the popular ones.

reliability coverage

The share of the tool catalog with recorded harness results. The live coverage label and tool count appear in the reliability totals above.

Reading the numbers honestly

Three claims the data supports

1. Every receipt in the sample is hash-backed and public. 93 of 93 carry a valid result_sha256 and a public receipt page.
2. Verification is continuous, not one-off. The corpus spans 9 days with calls on 6 of those days, and the reliability harness last ran on 2026-09-25 08:10:05 UTC.
3. Work traffic is half the platform. 15,445 of 30,710 all-time calls (50.3%) are work tool calls.

Three claims the data does not support

Chain-anchoring shares (the receipt schema has no anchoring fields), claim-vs-execution mismatch rates (no such field exists), and full-population statements (the corpus is 0.304% of all-time calls). This page does not make those claims.

Frequently asked questions