Anthropic stopped selling Intelligence and started selling Infrastructure 🤖💸; Revolut just launched AIR, its AI Assistant. PRAGMA is why that matters 🤖📲

FinTech is Eating the World, 9 April

Hey Everyone,

Good morning & happy Thursday! Today we’re diving deep into Anthropic, which just made the biggest infrastructure play in AI since AWS made servers boring (breaking down Claude Managed Agents, why the architecture is really a platform lock-in strategy disguised as a product launch, which startups just lost their reason to exist + how to solve AI Agent context window using Skill Graphs & deep dive into Anthropic’s leaked Claude Source insde), and Revolut that just launched AIR, its AI assistant, but the real story is PRAGMA, the proprietary model trained on 40 billion behavioral events that no competitor can replicate (why the actual competitive threat isn’t Starling or Monzo Bank but Apple & Google embedding finance into the OS + bonus deep dive into Revolut’s 2025 financials, how Stripe built AI but for payments, & how to make AI your Sr. Financial Analyst inside). So let’s just jump straight into the finnovative stuff 🌶️

Anthropic stopped selling Intelligence and started selling Infrastructure 🤖💸

The news 🗞️ The most important AI launch this week wasn’t a smarter model. It was plumbing 🪠

Anthropic’s Managed Agents, now in public beta, is a fully hosted runtime for autonomous AI agents. Not another wrapper around an API. A managed execution environment:

→ sandboxed Linux containers spun up on demand

→ persistent append-only session logs that survive crashes

→ a credentials vault that keeps your API keys and OAuth tokens outside the sandbox entirely

→ crash recovery via checkpoint-and-resume

→ end-to-end OpenTelemetry tracing, and multi-agent orchestration in research preview.

You define the agent - system prompt, tools, guardrails - in natural language, YAML, or code. Anthropic runs everything else 🤖

The architecture splits cleanly into three layers: a stateless “brain” (Claude plus the orchestration harness), disposable “hands” (sandboxes that can fail and be re-provisioned without data loss), and a durable session log that lives outside both. If the harness crashes, a new one calls wake(sessionId) and picks up where the last one stopped.

So let’s dive deep here and break down the exact pricing and unit economics, why the architecture is really a platform lock-in strategy disguised as a product launch, how the security model solves the specific procurement objection that has blocked enterprise agent adoption for two years, which startups just lost their reason to exist, and the 3 signals over the next 12 months that will tell you whether Managed Agents becomes the AWS of agentic AI or another overhyped beta.

We also name the job title that didn’t exist in 2024 and may be the fastest-growing role by 2027.