Editorial illustration: machine signals passing through an optical instrument. This is a conceptual image, not a diagram of the measured traffic.
183,291 requests. 4,066 task-shaped requests. More activity—but not proof of more useful work.
Our last dedicated Agentic Web Observatory report was published on September 18. Sixty days later, we have extracted a new fixed snapshot of what reached WagerX’s MCP and A2A gateway. The October 4–5 field reports covered outbound agent experiments; they were not updates to these inbound traffic measurements.
Observation window: September 19, 2026, 00:00 UTC to November 17, 2026, 00:00 UTC, end exclusive: 59 complete UTC days, through November 16. Extracted: November 17 at 11:18:15 UTC. Today’s partial day is excluded.
The main finding is not simply a larger request count. Across two equal 28-day windows, total requests increased 63.5%, while task-shaped requests increased 32.9%. Task-shaped traffic therefore fell from 2.47% to 2.01% of raw requests. That is a change in the observed request mix—not a measured conversion rate.
What reached the gateway?
| Request class | Requests | Share of raw requests |
|---|---|---|
| Discovery | 115,751 | 63.2% |
| Task-shaped | 4,066 | 2.2% |
| Other | 63,474 | 34.6% |
| Total | 183,291 | 100.0% |
Discovery covers initialization and capability-listing methods. Task-shaped means MCP tool-call methods and the specified A2A message/task-submission methods. Other includes notifications, monitoring and methods outside those definitions. These are protocol-method classes, not judgments about whether a request was useful.
Repeated discovery polls are preserved in these counts. 183,291 requests does not mean 183,291 agents, users or companies. The live Observatory also provides qualified figures that deduplicate discovery. Those rolling dashboard figures should not replace this report’s repeat-preserving snapshot.
Equal windows: more traffic, a smaller task-shaped share
Loading interactive graph. Exact figures remain available in the report tables.
| Metric | Sept 22–Oct 19: 28 days | Oct 20–Nov 16: 28 days |
|---|---|---|
| Total requests | 67,570 | 110,451 |
| Discovery | 45,433 | 66,686 |
| Task-shaped | 1,671 | 2,220 |
| Other | 20,466 | 41,545 |
| Task-shaped share of raw requests | 2.47% | 2.01% |
These periods are disjoint and equal in length. September 19–21 remain in the full release but are outside this comparison. We do not compare the 59-day total directly with September’s much shorter report as a growth rate.
There is also a recent countertrend. November 10–16 recorded 399 task-shaped requests, down from 488 on November 3–9, while total requests were nearly flat: 28,186 versus 28,003. The larger multiweek total should not hide that latest weekly decline.
Records are present on every day in the release. That establishes retained daily coverage, not proof of uninterrupted instrumentation. Changes in public exposure, routing, tools, testing or caller composition may affect the counts. This observational comparison does not identify what caused the change.
A2A carried most task-shaped requests
Of the 4,066 task-shaped requests, 3,623 used A2A and 443 used MCP. This describes the methods that reached this gateway, not the relative popularity of those protocols across the agentic web.
All 4,066 returned HTTP 200. Their stored outcomes were different:
| Recorded task-shaped outcome | Requests |
|---|---|
| Success | 3,965 |
| Rejected | 61 |
| Unknown tool | 38 |
| Invalid arguments | 2 |
A successful protocol response is not independent verification that an answer was correct or a useful task was completed. A structured rejection can be appropriate behaviour. These records do not establish betting transactions, deposits, affiliate conversions or completed commercial work.
What MCP callers asked the registered tools to do
The registered MCP capabilities accounted for 405 requests. The remaining 38 MCP tool-call requests had unknown-tool outcomes.
Loading interactive graph. Exact figures remain available in the report tables.
| Registered capability | Requests |
|---|---|
| Agentic statistics | 148 |
| List casinos | 54 |
| Check a casino | 46 |
| Find agentic services | 44 |
| Bonus information | 29 |
| Regulatory intelligence | 26 |
| Top casinos | 15 |
| Country recommendations | 10 |
| New casinos | 10 |
| Long-running casinos | 9 |
| Latest audits | 8 |
| Compare casinos | 6 |
The statistics tool was the most-called registered capability. That is evidence of requests for the Observatory’s own data, not proof that an external AI provider cited it or used it in a final answer. The country-recommendation count names a capability only; no caller geography or requested country is disclosed.
Known test labels are filtered, but not every internal or synthetic request can be identified. These capability counts must not be described as external-only demand, unique customers or organic adoption.
Suspicious payment language remains a separate measurement
Applying the previous report’s two case-insensitive text-pattern checks found 47 A2A task-shaped requests with payment-solicitation indicators. The matching rules are included in the fixed aggregate dataset.
These 47 requests already appear in the task-shaped total. They are not an extra category to add. Pattern matches are not verified scams or a count of attackers: broad terms can produce false positives, while encoded messages and different wording can evade the check.
No raw messages were retrieved into this report. We ran no new live replays and make no new claim about what those historical messages received or whether their contents were correctly recognised.
What builders can take from this release
For companies building agentic iGaming services, the useful distinction is between being discovered, receiving a task-shaped request and demonstrating a completed useful task. These records measure the first two at a protocol level. They do not independently establish the third.
Discovery still accounts for most requests. Information tools receive calls, including casino checks and regulatory lookups, but their counts remain much smaller than discovery traffic. A larger public endpoint footprint should not be reported as customer adoption without additional evidence.
The next measurement challenge is verified usefulness: did an authorised caller obtain a grounded answer that resolved its question? That would require additional privacy-conscious evaluation. It cannot be reconstructed from HTTP status codes or a success label alone.
Methodology, retention and reuse
This release uses read-only production aggregates from records created within the stated half-open UTC window, with a nonempty IP field. It sums COALESCE(request_count,1), preserving repeated discovery counters. It excludes case-sensitive user-agent labels beginning Werkzeug/, and labels beginning WagerX- that also contain Verification. Those exclusions do not identify all possible internal tests.
The previous release separately excluded three specifically labelled September 18 safety replays. September 18 is outside this release. We start on September 19 to avoid overlapping its partial final day; the remaining hours of September 18 are not covered by this new snapshot.
The cutoff is midnight UTC, a closed ten-minute discovery-counter boundary. No selected row had a completion timestamp at or beyond the cutoff. Daily counts, protocol splits, registered capability counts and definitions are preserved in the JSON.
The current gateway code prunes raw request records after 90 days. This matters for a longitudinal Observatory: dated aggregate releases preserve selected observations beyond raw-log retention. They do not preserve the private underlying records, and older snapshots must not silently be replaced with rolling values.
Comparable broad discovery logging began after the August 31 instrumentation expansion; September 1 was the first complete UTC day under that coverage. This report does not treat the earlier sparse August logs as the same adoption series.
Scope: observations at one WagerX gateway, not an industry census, market-share estimate or independently verified adoption study. No countries, raw prompts, IP addresses, user agents, caller identifiers, request-level records, confidential company information or bank/card details are published.
Report series: September 18 report · Download the November fixed snapshot · Live Observatory · Gateway documentation.