The Intake
Intake — May 6, 2026
Editor’s note
Two of today's items extend what an agent can do without a human in the chain — Cloudflare and Stripe by handing agents account creation and a payment token, Anthropic by shipping templates that take a finance analyst's week-of-work and put it under one approval gate. The third item is local-inference plumbing: Gemma 4's new MTP drafters, Apache 2.0, claiming up to 3x at single-request batch sizes.
On the substrate
Cloudflare and Stripe ship a provisioning protocol agents can run end-to-end
A coding agent can now create a Cloudflare account, start a paid subscription, register a domain, mint an API token, and deploy to production in a single session — no human steps required except accepting Cloudflare's terms of service. That is the capability Cloudflare published April 30, built on a three-part protocol co-designed with Stripe: discovery via the Stripe CLI's catalog of provisionable services; authorization, in which Stripe attests user identity and provisions a Cloudflare account if none exists; and payment, with a default cap of $100 per month per provider that the user can adjust. Cloudflare frames the protocol as an open standard — any platform with signed-in users can expose its services through the same OAuth, OIDC, and tokenization primitives. The change is at the account-and-credential acquisition step, which previously required a human at a login screen.
Anthropic ships ten finance agent templates and the M365 add-ins to land them
Anthropic The Register Bloomberg
Ten prebuilt finance agent templates landed May 5 from Anthropic, split five-and-five: pitch builder, meeting preparer, earnings reviewer, model builder, and market researcher on the research-and-coverage side; valuation reviewer, general ledger reconciler, month-end closer, statement auditor, and KYC screener on the finance-and-operations side. Each template ships as a plugin for Claude Cowork and Claude Code, and as a cookbook for Claude Managed Agents — the cookbook path adds long-running sessions, per-tool permissions, managed credential vaults, and a Console audit log. Claude Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37%. Vals AI is independently operated. Add-ins now cover Excel, PowerPoint, Word, and Outlook, and the data-connector list adds Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, and Guidepoint to the existing FactSet, S&P Capital IQ, MSCI, PitchBook, and Morningstar coverage. Anthropic's framing is that users review and approve before anything is filed, sent, or acted on — language pitched at compliance teams.
For operators
Gemma 4 MTP drafters claim up to 3x with no quality cost
Google AIToolly claypier llama.cpp PR #22673
Google released MTP drafters for the full Gemma 4 family May 5 under Apache 2.0. The architecture pairs a lightweight drafter with the target model: the drafter predicts several tokens in parallel using the target's activations and shared KV cache, and the target verifies in one forward pass and accepts or rejects against its own distribution. Google's reported numbers are up to 3x at single-request batch sizes and roughly 2.2x for the 26B MoE variant on Apple Silicon at batch sizes of 4 to 8. Compatible runtimes already include Hugging Face Transformers, MLX, vLLM, SGLang, and Ollama; llama.cpp support is in beta and tracked at PR #22673. Weights are on Hugging Face and Kaggle. If Gemma 4 is running locally in your stack today, the drafter is the cheap action this week.