Week of September 13 – September 19, 2026
The Agent Bill Arrived, and It Is Rewriting the Architecture Argument
For three weeks this newsletter has been tracking an argument about control — who the agent is, what it is allowed to do, and whether anyone is checking. That argument is still running and I will get to it.
But something else happened this week, and I think it matters more, because it is the first time the agent conversation has been settled by arithmetic rather than by opinion.
Redocly published a measurement: letting the model write code against your API, instead of chaining MCP tool calls through the context window, cuts LLM costs by up to eighty percent. Vonage published a second one from a completely different angle — the JSON Schemas you ship in your tool definitions carry a token tax, paid on every single call, that almost nobody has costed. And cData put a number on the spread between models doing the same job: 178x.
Two weeks ago Jens Neuse told me on the podcast that past about five MCP tool calls you are gluing them together with inference, and inference is slow, expensive and unpredictable, so the architecture is simply wrong. I wrote at the time that models are good at writing code and we should let them. It was an architectural argument and you could reasonably have disagreed with it.
It is not an architectural argument any more. It is a line item.
That is a different kind of pressure, and it moves faster. Nobody rewrites a working integration because a specification is more elegant. Plenty of people rewrite one when somebody shows them the bill.
The Agent Bill Arrived
- The headline measurement: Redocly’s code mode for MCP, up to 80% lower LLM costs. Read it next to Vonage on the hidden token tax in your JSON Schemas, which is the best-titled post of the week and the one I would hand to anyone who thinks the description layer is free. Your schema is not documentation any more. It is payload, and you are billed for it on every call.
- cData measured a 178x token cost spread between models on the same job and asked whether your cheapest model can do it. That question was unaskable a year ago because nobody had the comparison.
- A whole FinOps layer formed around this in one week. Cast AI on attributing inference spend to teams on Kubernetes. Finout on AI unit economics and allocating token spend to the teams driving it. Larridin on tracking token costs across multi-agent workflows. When cost-allocation vendors show up, the spend has become somebody’s budget line rather than somebody’s experiment.
- And the context window itself became a design constraint people write about seriously: DZone says stop overfeeding your agent’s context window, while DataHub argues the hard part is not the context window at all, it is the context engineering. Those two are less contradictory than they look. One is about volume, the other about curation, and the bill punishes you for getting either wrong.
Here is the part I want to flag, because it cuts against the last month of this newsletter. Everything above rewards a smaller surface — fewer tools, thinner schemas, less context, code instead of chained calls. Everything in the section below rewards a richer one — identity, policy, audit, provenance, governance metadata attached to every call. Those two pressures are now pulling on the same protocol from opposite ends, and I do not think anyone has noticed they are in tension yet.
Meta Shipped MCP for WhatsApp Business, and the Agents Still Have No Identity
- The launch: Meta announced Business Messaging MCP — set up and manage WhatsApp Business from your AI agent. That is one of the largest messaging surfaces on earth becoming agent-configurable.
- The catch, and it is the same catch as the last three weeks: Meta lets Claude and Codex configure WhatsApp Business via MCP, but the agents do not get their own identity. The agent acts as you. Every action it takes is indistinguishable, in the log, from an action you took. I have now written some version of this sentence four weeks running, and the pattern is not that vendors are careless — it is that shipping the capability is a quarter’s work and shipping the identity model is somebody else’s roadmap.
- The vendors kept shipping the identity model anyway. Auth0 launched Agent Gateway, an identity control plane for customer-facing AI — in beta, and note it is a gateway, which puts it in the same rack as the three other gateway categories I flagged a fortnight ago. WorkOS extracted its Atlas experiment into Airlock, a shipped agent governance product. JumpCloud argued the agentic era demands a unified identity strategy.
- The sharpest thinking came from Traefik, twice. The two questions behind every agent authorization standard is the clearest framing of this problem I have read, and agentic telemetry is not an audit trail names a confusion I see constantly — that having logs is the same as having accountability. It is not. A trace tells you what happened. An audit trail tells you who authorized it.
- Cequence, consistent with what they said last week, argues agentic AI requires governing actions and behavior rather than identity. And Palo Alto’s Unit 42 went and looked at AgentCore harness credentials — the uncomfortable space between a vault and a heap you can read. That is the research end of the same problem, and it is the kind of work that eventually forces the roadmap.
📄 Fundamentals of Context Engineering — If the first section of this issue is right, then what you put in front of a model is now a cost centre, not a convenience. Agents do not magically understand your API. Somebody has to curate what they need, decide what to leave out, and know which of your existing artifacts already carry that weight. This is the practical version: what context an agent actually requires from a provider, where it should come from, and how to stop paying a token tax on material the model never needed. $25
One Company Published Twenty-Three Help Pages, and the Real Count Went Down
I count 86 third-party posts naming MCP in a title this week, up from 77. That looks like acceleration. It is not, and this is the kind of number I would rather correct in advance than have you quote.
Twenty-three of those 86 are Botify, published across two days — an overview, FAQs, troubleshooting, example prompts, a migration guide, and sixteen individual use-case pages. Take that one help centre out and the count is 63, down from 77 the week before.
So the honest read is: MCP publishing fell this week, and one company’s documentation sprint disguised it.
- But Botify doing that is itself the story, and it is the arc I flagged three weeks ago completing. MCP documentation is now a shipped product deliverable with its own information architecture — overview, FAQ, troubleshooting, use cases, migration — not a launch post. Drata shipped a setup and usage guide. Thoropass documented connecting to its server. Frontify published use cases and prompts. Knak published five separate integration pages in a week.
- And the tell I have been waiting for: Rippling MCP is down. A status-page incident. Not a blog post, not a comparison, not a framework — an outage, on the status page, like any other production dependency. You do not get an incident page until customers are paged when you break.
MCP Grew a Price List and Lost Its State
- Zuplo laid out four MCP monetization models for agentic commerce. A protocol gets monetization models at the point where somebody’s MCP server is a cost centre they would rather make a revenue line. Put that next to the first section of this issue and the shape is obvious: the calls cost money, so somebody is going to charge for them.
- WorkOS explained what changed when MCP went stateless in the 2026-07-28 spec, and separately made the case for governing tool calls beyond OAuth. The stateless turn is underdiscussed and it matters for everything above — stateless is cheaper and more scalable, and it also means the server remembers nothing about what the agent already did.
- The enterprise end kept arriving: GitLab shipped MCP tools for platform teams to scale automation safely, AWS wrote up defense-in-depth authorization for MCP tools on Amazon Quick, and Solana Explorer launched a server for on-chain data.
- Two good sceptical reads to end on. Postman’s SDK vs CLI vs MCP, a dev story walks the same decision the cost section is now forcing. And DoltHub’s “So you want an MCP Database?” is the kind of post that asks whether the obvious thing is actually a good idea, which almost nobody writes during a boom.
From My Desk: A Week Spent Inside the Specifications
I spent this week almost entirely in the specification layer, and I should be straight with you about the contrast. Across 7,556 third-party posts, OpenAPI was named in exactly one title — Redocly’s, on generating OpenAPI from real traffic. That is three weeks running at one. The layer everything above depends on remains almost entirely undiscussed by the people depending on it.
- Arazzo had a real week, and it is the most active I have seen that specification. Arazzo is adding SOAP, and the pull request wants comment before the release. Arazzo wants to pause a workflow and wait for you — which is elicitation, arriving in the workflow specification rather than the protocol, and worth reading next to the MCP version. One review turned the Arazzo gRPC pull request into an RPC pull request, which is governance working exactly as it should, in public, in a day. And Big City Storytelling: Arazzo.
- The vocabulary thread turned into findings. Sixty-two companies declared an OpenAPI in Agent Resource Discovery and spelled it eleven different ways. Windmill produced the first real collision in the MCP vocabulary. Zoho quietly shipped the largest MCP vocabulary deployment anywhere. Wistia built the best agent vocabulary nobody knows about. Eleven spellings of one word is what an unregistered namespace looks like from the inside.
- On specifications themselves: I mapped how nine API specifications point at each other, which is the piece I would point at first. API specifications are living. APIs.json is an anthropological specification. VS Code is finally catching up with JSON Schema 2020-12.
- Elsewhere: Onboarding shipped as an agent skill. Is OpenAPI still relevant when you tell Claude “make me an API”? Fern puts SDK design decisions in the contract and Mintlify does a lot with one OpenAPI extension. Federation aligns schemas, but it cannot align people. The business and politics of API portals and gateways. And Microsoft works in the open with 70,000 engineers — you can publish your API documentation.
No conversation this issue — the second week without one, and I will have new ones shortly.
I want to end where I started, because the two halves of this issue are the same story told from opposite ends. The cost people are about to tell you to send the model less. The governance people are about to tell you to attach more to every call. Both are right. Nobody is going to reconcile them for you, and the place that reconciliation has to happen is the description layer — the thing that can say, once, what an operation is and what it needs, so the runtime does not have to re-explain it on every request and pay for the privilege.
That layer got named in one title out of seven and a half thousand this week.
See you next week.
🔌 APIs.io — what every provider charges, free to read — This issue is about the bill, so here is the part of the catalog that speaks to it, and the honest framing is that most of it costs you nothing.
I hold 9,429 pricing-plan profiles, 9,680 rate-limit profiles and 6,875 FOCUS-aligned FinOps profiles across the catalog, and reading any of them is free. No key, no tier — cardinality is the line, so everything about one provider stays open.
The operation worth your week: before you commit to a provider, read what it actually charges and what it actually limits, rather than the marketing page.
curl "https://apis.io/api/v1/plans?limit=50"
curl "https://apis.io/api/v1/finops?limit=50"
Same data as the find_plans and find_finops MCP tools. What is not free is ranking or comparing across the set — leaderboards, cohorts and the scored dataset are Understanding, $199/mo. There is also a per-call Access tier at $0.01 in the plans table; it is listed as not yet purchasable, so treat it as a preview rather than something to budget for.
