APIdays Paris: catalog refresh and suggestions for the three talks
Fresh look at the catalog on 2026-11-11, for refining the Paris decks. The rubric snapshot was re-taken at 0.27.0, with pages scored 2026-11-06: 38,836 providers, 9 facets, 22 agent-readiness dimensions (153 raw points). Earlier snapshots were 0.19.0 / 26,755 (London) and 0.23.0 / 28,597 (Nordic). _rubric/snapshot_rubric.py --check passes. It needed --scoring pointed at master’s scoring.yml, because the api-search/signals checkout sits on a feature branch at 0.26.0.
What was done
| Deck |
Cloned from |
Refreshed |
2026-apidays-paris-inheritance/2026-apidays-paris-inheritance-outside.html: Agents Inherit Your Governance |
London inheritance deck |
Title, venue, dates. All 32 rubric slides swapped for 0.27.0, plus the 3 new dimension slides. Provider counts in the frame. |
2026-apidays-paris-bolt-on/2026-apidays-paris-bolt-on-agent-hype.html: You Cannot Bolt Agent Readiness Onto Ungoverned APIs |
Nordic agent-hype deck |
Title (was “The Agent Hype is at Odds…”), venue, dates. 13 rubric slides swapped for 0.27.0. Provider counts. |
2026-apidays-paris-eu-gov-keynote/: Keynote: What Europe’s Government APIs Look Like From the Outside-In |
new, Munich reveal.js template |
— |
New commentary for the three 0.27.0 dimensions (Pagination, Distinct Operations, MCP Tool Quality) is in _rubric/commentary.yml. Regenerating the core rewrote every deck’s generated *-rubric-core.*. No hand-authored deck other than the two clones was touched.
Still stale, for you to refine:
- Speaker notes and HTML comments in both clones still say London, Stockholm or Toronto.
- The Nordic deck’s hand-authored MCP finding slide (“3,347 providers with an MCP server, 1,266 reach Agent-Native…”, measured 2026-10-03) was not regenerated. Re-measure it before presenting.
- The bolt-on deck’s frame slides still argue from the Nordic room (PSD2, GDPR, DORA framing). Most of that travels to Paris, but the room is different.
How the catalog moved (share of providers earning each dimension)
| Dimension |
Sep (0.19) |
Nov (0.27) |
|
| MCP Server |
8.6% |
16.4% |
nearly doubled |
| Well-Known Catalog |
1.8% |
8.9% |
×5. See the caution below. |
| Protected Resource Metadata |
2.2% |
10.6% |
×5. See the caution below. |
| Registration Without a Human |
3.1% |
8.4% |
|
| Delegated User Identity |
7.6% |
13.4% |
|
| A2A Agent Card |
0.4% |
2.6% |
|
| Idempotency |
3.4% |
4.7% |
|
| Machine-Readable Contract |
34.0% |
32.7% |
flat (12k thinner providers joined) |
| Stable Error Semantics |
20.6% |
22.5% |
flat |
| Rate-Limit Signaling |
35.8% |
30.5% |
down |
| new: Distinct Operations |
— |
23.8% |
|
| new: Pagination |
— |
6.4% |
|
| new: MCP Tool Quality |
— |
1.4% |
|
- Agent readiness above the composite: 3.3% of providers in September, 6.1% now. Composite × agent correlation went from r = 0.882 to 0.869.
Suggestions for evolving the talks
- Lead with the divergence. It is the strongest new evidence for both talks. Agent-surface signals (MCP, well-known catalogs, registration, delegated identity) roughly doubled to quintupled in two months, while the foundations (contracts, errors, rate limits) are flat or down. That is “you cannot bolt agent readiness on” happening in the data, in real time. Twice as many providers now score higher on agent readiness than on their foundation.
- But check the bolt-ons are real before saying so. Some of the quintupling is measurement, not market. Open issues: #1138 (well-known catalog credited where only an HTML shell or a docs platform’s file answers) and #1061 (protected-resource metadata credited from another organisation’s host). Run the refreshed numbers again after those fixes land. If the growth survives, the claim is stronger. If it shrinks, that is also a story: agent signals are easy to fake, even by accident.
- Use the new dimensions as the “inheritance” proof. Distinct Operations (23.8%) and Pagination (6.4%) are pure governance-of-the-contract checks that agents inherit directly. A tool picker can’t choose between duplicate summaries, and an agent that can’t page truncates silently. MCP Tool Quality (1.4%) is the bolt-on version of the same problem: an MCP server whose tools are vague is governance debt in a new wrapper.
- Bring Europe into both rooms. Same week, same venue. Point to the keynote’s evidence: 0 of 16 agent runs found a government API through a well-known catalog, and the national design rules (NL ADR) show up as real contracts. For bolt-on, Italy is the cautionary tale (a strong platform the outside can’t see). For inheritance, the Netherlands shows governance inherited body by body.
- Tie both to the OpenLint launch. Governance rules are what agents inherit. Seven European public bodies already write rulesets separately (openlint.org/blog, 2026-11-11). One shared slide across the three talks keeps the message consistent.
- Re-freeze once. Pick a freeze date after #1138 and #1140 land, re-run
snapshot_rubric.py → build_slides.py → assemble_decks.py, then re-swap the rubric slides in the hand-authored decks. The swap script logic is in this session’s notes: match by slide class and <h2> title.