Week of October 18 – October 24, 2026
Everyone Is Selling Governance, and the Door Is Where It Lives
I was in New York this week, opening Tyk’s LEAP ON TOUR: The Future of MCP Governance for a room of about thirty people who decide how their companies will govern agents, most of them from finance. My keynote was Discovery Is the First Governance Control: you cannot govern an agent’s access to something you do not know you have. Afterward I sat on a panel that kept circling one question, who owns the agent, and Tyk wrote up what New York taught them about MCP governance — three owners, in practice, means nobody is accountable.
Then I came home to the blog pull and found the whole industry having the same conversation. I read 12,596 posts from 2,758 provider blogs this week, against 13,039 from 2,828 the week before. “Governance” appeared in 48 titles. Some of that is content marketing, but a lot of it was product: Google, GitLab, JetBrains, Traefik, Saviynt and Workato each shipped or positioned a way to govern what agents do.
Everyone Is Selling Governance, and the Door Is Where It Lives
The argument that clarified the week came from Traefik, in two posts. Harnesses and sandboxes will keep multiplying; your governance should not: every team adds a harness with its own rules for reach, credentials, cost and approval, and what every agent shares is the way out. Govern at the door, not in the sandbox, using NVIDIA’s OpenShell as the example. That is the gateway argument this newsletter has been making for a month, now made about agents instead of APIs.
- Google introduced Agent Compute Engine and expanded agent governance, turning the infrastructure behind Gemini’s agents into a cloud service.
- GitLab shipped an AI governance dashboard with one view of agent activity — its own agents and the outside ones, Claude Code, Cursor and Codex, in one place. JetBrains made Air Governance generally available to manage which agents, models and providers an organization uses, and what it spends. Saviynt introduced action governance for agents inside business-critical applications.
- And the sentence I would argue with: Workato says AI governance is the future of API management — API management controlled who could call an endpoint; AI governance controls what an agent does once it gets there. I agree with the distinction. I do not agree that the second replaces the first. You cannot govern what an agent does at an endpoint you never inventoried.
That is the point I made in New York, and the reason I keep writing about the unglamorous layer underneath. Bruno Pedro argued well that API governance isn’t just rule automation, and I answered that most teams haven’t automated the rules yet — which matters, because a governance dashboard can only show you the agents and APIs that something already knows about.
The AI API Layer Is Consolidating
- OpenAI donated Open Responses, its specification for a common LLM API, to the Agentic AI Foundation — the Linux Foundation home that already holds MCP, A2A and AGENTS.md. I wrote that Open Responses gets a neutral home, and a baseline. The foundation is quietly becoming where the agent era’s contracts live, and Vultr joined it as a Gold member the same week.
- Decision models kept becoming ordinary API calls. OpenAI published the launch post for its Decisions API — fast, focused, low-cost decisions for any application. Zoom showed its Decisions API routing customer complaints, Snowflake brought Decision Fast to Cortex AI over SQL or REST, SAP made RPT the decision model in Joule Studio 2.0, and LiteLLM previewed JevRank, decision models behind its rerank API. Two weeks ago I counted eleven request shapes for the same three questions. Every one of these adds a twelfth.
- I spent the week on what all of this does to the answer itself. DeepSeek open-sourced DeepInfer, and batch-invariant inference, deterministic sampling and schema-constrained output move the hard problem to the shape of the answer. Then I walked that through three regulated industries: healthcare, payments and banking, and insurance, where the shape of the answer is behind a login.
From the research: The Fundamentals of API Discovery
You can’t govern, secure, or automate what you can’t find. After fifteen years working on API discovery, I wrote down the fundamentals — why two decades of directories never solved it, why inventory is archaeology not a query, and where the old human problem becomes urgent at machine scale: discovery for agents.
The Missing Piece Is Proof of Authority
Agent identity and agentic commerce spent the week converging on the same gap.
- Auth0 expanded its Client ID Metadata Document support — pre-registration and dynamic registration — and paired it with application profiles to manage agent clients at scale. That is the standards plumbing of agent identity arriving in a mainstream identity product. Auth0 also laid out one identity layer across the four channels agents now shop through.
- The phrase of the week came out of Money20/20: agentic commerce is here, and the proof of authority is missing, from a panel with Visa Intelligent Commerce, Circle, TD Bank and Proof on how banks authorize agent payments. Cover Genius made the matching argument: agentic commerce has a payment layer, and now it needs a protection layer.
- The rail kept growing anyway. Algorand announced the winners of its Global x402 Challenge: 115 projects put a paid x402 endpoint live on mainnet and settled real USDC.
My own work this week sat right in that gap. The human in the loop is now one email — the follow-up to the agent that got four APIs to trust it, closing the part that was still broken: getting a person’s authorization in front of a provider. And responding to Curity’s Mike Goerlich: trust you can carry has to be written down.
Counted, Not Charged
Monetization came back into the conversation, mostly as a question nobody has finished answering.
- Workato’s story about DMSi, a 350-person ERP vendor, monetizing its agentic tier is the most honest one: build one controlled access point and the guardrails first, then work out the price. Thales wrote about the move from seats to usage, and Salesforce about tiered monetization on AgentExchange.
- My two cents, literally. I came out of Josh Twist’s talk in Stockholm arguing that API monetization is falling flat, and maybe it is the word. Then I measured it: gateways count every token and charge for almost none of them. The metering already exists. The decision to bill on it mostly does not.
From My Desk: Eighty-Six Markets, and a Maturity Model We Left on the Table
- The Market layer grew fast. A week ago the Kin Score’s market dimension covered four markets. It now covers eighty-six, and every regulated market already publishes its own map. Along the way: the brands are missing from their own markets, a listing is not a certification, five thousand official SDKs and nine thousand that are not, and 389 agent cards are live. Here is where the Market layer stands, warts and all.
- We had the agent protocol in 2008. Leonard Richardson’s maturity heuristic topped out at hypermedia, and we left it on the table. The follow-up is the one I would read: we did not skip hypermedia, we moved it out of the response — into the catalogs, cards and descriptors this newsletter writes about every week. Mike Amundsen’s two axioms for designing for autonomy belong in the same conversation.
- Harnesses. Naresh Jain’s measurements of coding-agent harnesses are the most useful thing I read on agents this year, and the harness that works is a contract. I also built a competitive landscape harness for a provider standing up its own MCP server.
- MCP in the field. A banker in New York told me finance does not have enough useful MCP servers, so I asked the catalog, and they are right about the banks. Meanwhile only four companies serve skills over MCP the way the specification says, and the first Poppy files have appeared, including one a vendor wrote.
- And the rest. 380 companies say Claude and 142 say Anthropic — job postings name products, not companies. The Incumbent and the Challenger: NCR Voyix and Lightspeed.
One ask, and it is not for me. If you depend on OpenAPI, please become a member of the OpenAPI Initiative, and if your membership lapsed, come back. Every governance product in the first section of this issue reads a contract somebody had to standardize, and the people who maintain that standard do it on very little. Across 12,596 posts this week, OpenAPI was named in four titles — one of them APIAddicts deprecating its Sonar projects for OpenAPI and AsyncAPI, which is what unfunded tooling looks like when it runs out.
On the numbers: I count 137 titles naming MCP this week, against 226 last week. Nearly all of that drop is one MCP directory, vinkius.com, which published a hundred listings the week before and none this week. Without the directories it is 84 against 88 — flat.
See you next week.
🔌 APIs.io — find what you have before you govern it — My keynote this week was that discovery is the first governance control, so here is the discovery. APIs.io profiles tens of thousands of API providers — their contracts, MCP servers, agent cards, rules and markets — and searching it is free. If you run an API program, start by looking yourself up: what the outside world can see of you is the inventory every agent is already working from. And if you want the same thing pointed inward at your own estate, that is APIs.io On-Premise.
Also on the shelf: The Fundamentals of API Evangelism
Your developer blog is the most underrated asset you have — it compounds. Every post is a door that keeps letting people in years after you wrote it. That’s one thread in my fundamentals of API evangelism paper.
