Week of November 8 – November 14, 2026
The Agent Carries Your Token Now, and the Skills It Installs Are a Supply Chain
By Kin Lane · Sent November 16, 2026
I read 15,034 posts from 3,278 provider blogs this week, up from 14,429 from 3,277. The number of blogs held steady while posts rose, which means the same companies wrote more, and a good part of what they wrote was about agents.
Last week the agent got a login of its own. This week the question moved one step further: when an agent acts for a person, whose authority is it carrying, and what is it installing along the way?

The Agent Carries Your Token Now
- The platforms shipped the standard way to hand authority down. Amazon Cognito now supports OAuth 2.0 token exchange (RFC 8693), so an agent acting on a user’s behalf can swap the user’s token for a narrower one. AgentCore Identity added OAuth 2.0 discovery URLs, WorkSpaces for AI Agents added OAuth, and Azure Monitor’s Grafana MCP dashboards run “signed in as you”.
- The practitioners wrote down what has to travel with the work. Dome Systems published a checklist for carrying delegated identity into a queued task: the subject, the actor, the scope and the expiry. That same week it argued that you should use rate limits to protect a backend, not to decide authority. Trigger.dev explained how OIDC workload identity works.
- The security vendors pushed back on the idea that an agent is one thing. Nudge Security: an agent isn’t one identity, because every MCP and API connection borrows its own credential, often a person’s OAuth grant. C1 called putting AI to work an identity problem, and Mitiga’s ConsentFix write-up shows OAuth consent phishing getting past MFA in Entra ID.
- And the friction is real. DocSpring explained why Claude Code MCP servers keep asking you to sign in again.
Token exchange is the right primitive, and it is good to see it arrive as a checkbox in a large identity provider instead of a custom build. But an exchanged token is only as narrow as the scopes the API publishes. If your API has one scope called read and one called write, the agent’s token will be exactly that broad, however carefully it was exchanged.

The Skills Agents Install Are a Supply Chain
- The numbers came in from several directions at once. More than 575 malicious agent skills shipped trojans, miners and stealers while posing as utilities. Mitiga scanned more than 50,000 AI instruction files in 7,000 repos and found hijacked base URLs and more than 1,230 leaking API keys. Knostic found hidden data collection in two MCP packages, and asked what should happen when an MCP tool asks for your GitHub token.
- The attackers are already looking. Honeylabs analyzed 366,845 requests from a scanner that asked every web server it reached for 32 AI coding assistant files, including OAuth token files and MCP server definitions. Swytchcode asked its own agent whether it was leaking API keys, and it printed one into the chat. Coinigy told customers to revoke the exchange API keys they had connected.
- The response is starting to look like package management. Cloudsmith wants to be the trusted control plane for agent skills. Stacklok argued that skills need the same governance as MCP servers. At the edge, Buildkite lets you choose read-only access when connecting its MCP server, and lemlist lets you control each MCP tool’s permissions.
We have been here before, with npm, PyPI and Docker Hub, and we know what works: signed artifacts, a registry you trust, and a lint pass before anything gets installed. A skill is a Markdown file that tells an agent what to do with your credentials. It deserves at least the scrutiny we give a dependency, and probably more.

From the research: The Fundamentals of the API Lifecycle
The protocol, the architecture, and the wave you rode in on are not what decide whether an API succeeds. Managing the whole lifecycle, end to end, is — threaded by one machine-readable definition.
The Protocols Got a Release Cycle
- MCP announced a more predictable release cycle, and opened the file upload and download discussion that, as The New Stack put it, developers have been working around for two years. GraphQL announced the GraphQL over HTTP specification, eleven years after the first release left transport out of scope.
- Providers versioned on a calendar. Qlik moved to quarterly API versioning with its 2026-11 release, and Shopify is deprecating a Storefront MCP tool in favor of the UCP MCP Ask tool, so MCP tools now get deprecation notices like any other API. Brave Search now takes agent payments over x402, settled per query in USDC, with no account and no key.
- The best measurement of the week: monit.rs found that four public APIs have not changed a byte since July, while GitLab shipped 12 spec revisions without breaking anything. Surface size predicts almost nothing about stability. And Snap Render asked coding agents to add screenshots 120 times. Not one picked a screenshot API, and every run that wrote code self-hosted Playwright.
- More servers reached GA and production: Retool, Red Hat Enterprise Linux, EasyPost, HubSpot, Clerk and Semgrep. On the design side, FrontMCP argued your OpenAPI doesn’t need 19 MCP tools, it needs four, and Sentry answered “MCP or CLI?” with yes.
I count 209 titles naming MCP this week, including the directories, up from 141. The Snap Render study is the one I keep thinking about. If an agent never picks your API when it builds something, it does not matter how good the API is. Being found by agents is now its own problem, separate from being good.

From My Desk: Agent Cards, Estates, and What a Score Is Worth
- Who publishes A2A agent cards. 934 conformant A2A agent cards, and who is publishing them. Then, one industry at a time: SAP published an A2A agent, and how Oracle and Salesforce compare, Siemens runs its own, and what I found at Schneider Electric and Rockwell Automation, and Booking.com announces its MCP server in an A2A card.
- API estates. The Amazon API estate: 199 surfaces, one source of truth, and Amazon and Microsoft: two API estates, two ways to build one.
- Fitting this week’s theme. Lint an agent skill before you hand it to an agent, the agent hype is at odds with your API foundation, and what our agents do when they profile an API provider.
- Specifications and standards. How the specifications relate, twelve OpenAPI validators and the rules none of them check, find the deprecated API operations that never say when they go away, and the business terms of an API, made machine-readable.
- What a score is worth. Benchmark your APIs against competitors you did not know were looking, what an acquisition target claims and what the score says, and what to look for in a vendor before you sign or renew.
- And the milestone. APIs.io now searches 41,300 providers, 203,000 APIs and over a million operations.
What You Can Get Done With These APIs
These Provider Stories are written for the person searching for how to get something done, who may not know an API exists. A few from this week:
- Add Stock And Crypto Investing To Your App With Alpaca
- Track Shipments And Send Delivery Notifications Automatically With FedEx
- Send Invoices, Pay Bills And Pull Financial Reports Automatically With QuickBooks
- Manage Users, Security Alerts And Team Plans In Microsoft 365 With Microsoft Graph
- Add Fast Site Search And Build AI Shopping Agents With Algolia
- Find And Fix Security Vulnerabilities Across Your Code And Cloud With Snyk
All three sections this week come back to the same thing. An exchanged token is only as narrow as the scopes the API publishes. A skill is only as safe as what we check before installing it. And an agent only picks an API it can find. The API’s contract, its scopes and its listing are each something a provider can publish this week.
See you next week.
🔌 APIs.io — see whether an agent can find you — Agents are choosing APIs on their own now, and often choosing none at all. Search APIs.io for your company and see whether your scopes, your contract and your pricing are described anywhere a machine can read them.
Also on the shelf: MCP Documentation Beyond Reference
An agent does not fail on your parameter types. It fails on the third call, in a state your docs never described.
