Cloudflare, Stripe & Anthropic Shape Autonomous AI

Welcome to Memorandum Deep Dives. In this series, we go beyond the headlines to examine the decisions shaping our digital future. 🗞️

This week, we look at how Cloudflare, Stripe, and Anthropic, in the span of a single week, introduced infrastructure components that grant AI agents persistent memory, economic agency, and operational autonomy, a convergence already accelerating enterprise demand as Anthropic's annualized revenue jumped from roughly $9B to $30B in four months.

The surface reading is straightforward: three companies shipped useful but incremental developer tools. That reading misses what these announcements actually mean for the future of AI agents that can now spend money, deploy software, accumulate institutional knowledge, and coordinate work across specialized systems without continuous human oversight.

The real story then is that once identity, payments, memory, and orchestration connect into a unified stack, the competitive moat shifts from model quality to the operational ecosystem surrounding it, and that distinction reshapes who captures value in the AI economy.

How AI got economic agency

For most of AI’s short history, artificial intelligence systems have functioned primarily as tools that depend on humans for direction. They could summarize reports, write code, and analyze information, but they still relied on people to actually execute tasks in the real world. Even the most advanced AI systems lacked the core attributes needed for operational independence: persistent memory across sessions, the ability to spend money, verifiable identity, and direct access to the infrastructure needed to deploy and manage what they created.

Within one week, Cloudflare, Stripe, and Anthropic introduced components of an infrastructure stack that fundamentally expand what AI agents can do.

Individually, none of the announcements looked revolutionary. Still, together they point toward a future in which AI agents no longer merely generate information but increasingly operate as autonomous participants in digital economies.

On April 30, Cloudflare announced that AI agents can now create Cloudflare accounts, purchase domain names, subscribe to paid services, and deploy applications to production environments without human intervention. The system, developed as part of Stripe Projects, combines OAuth-based identity delegation, tokenized payments, and machine-readable service catalogs. Together, they enable agents to discover services, authenticate themselves, authorize spending, and provision infrastructure autonomously.

The significance of this development lies less in any single capability than in how these systems now connect. Infrastructure-as-code tools have existed for years, and APIs have long enabled automated payments and provisioning. However, developers still had to custom-build the layers connecting them, while humans continued to handle approvals, credentials, billing, and deployment decisions. Stripe Projects reduces much of that friction by turning those fragmented processes into a more unified, agent-accessible workflow.

Giving AI economic agency

Under the framework, users approve payment methods and terms of service once, after which agents can transact within predefined limits, currently capped at $100 per month per provider, while keeping raw payment credentials hidden from agents.

In effect, AI systems are being granted constrained forms of economic agency. That distinction matters because once an AI agent can spend money, acquire infrastructure, and deploy software independently, it stops behaving like a passive assistant and starts functioning more like a junior operator inside an organization. The system can identify a need, acquire the necessary resources, and execute a workflow without waiting for continuous human intervention.

Teaching AI to remember

Just one week after Cloudflare’s announcement, Anthropic addressed another major limitation preventing AI agents from becoming operationally autonomous: memory.

At its Code with Claude conference in San Francisco on May 6, Anthropic introduced ‘dreaming’ for Claude Managed Agents. The capability enables agents to periodically review past work sessions, identify recurring patterns, extract successful workflows, consolidate lessons learned, and rewrite their memory stores into reusable operational playbooks that future sessions can reference.

Crucially, Anthropic is not retraining the underlying model itself. Instead, the system generates structured memory artifacts and plain-text notes that remain observable and auditable. Developers can require human approval before memory changes are accepted or allow the process to occur automatically. The result is something closer to institutional memory than human memory.

The rise of digital labor

Most AI systems today repeatedly relearn the same lessons because sessions remain isolated from one another. A coding agent may solve a deployment issue perfectly one day, then fail at the same task later because the operational knowledge was never consolidated into persistent workflows. Dreaming changes that dynamic by allowing agents to accumulate procedural knowledge over time.

Anthropic also introduced two features alongside dreaming: Outcomes, which let AI agents evaluate their own work against set standards, and multi-agent orchestration, which allows one agent to assign tasks to multiple specialized agents simultaneously.

Together, these systems begin to resemble a primitive operating system for autonomous digital labor.

An AI agent can now receive an objective, decompose it into subtasks, assign those tasks across multiple specialized agents, evaluate the outputs, consolidate lessons into long-term memory, and improve future performance iteratively. Combined with Stripe and Cloudflare’s infrastructure stack, the result is an increasingly complete operational loop.

The implications for software development are profound, as many functions that once required separate teams are increasingly being folded into software itself. Traditionally, engineers wrote code, DevOps teams managed infrastructure, finance departments handled payments and vendors, QA teams tested outputs, and operations teams documented workflows. The emerging agentic stack compresses much of that process into autonomous systems.

An AI agent can already generate functional applications. Now it can also purchase domains, provision infrastructure, deploy code, evaluate results, and improve future workflows based on accumulated experience. The gap between generating an idea and shipping a working product is shrinking rapidly.

That dynamic helps explain why at Anthropic’s Code with Claude conference, Mercado Libre was cited as running Claude Code across 23,000 engineers with a target of 90% autonomous coding by Q3 2026, a figure that describes a world where the majority of software engineering output at a major technology company is machine-generated, deployed, and iteratively improved by agents drawing on consolidated memory.

Governance falls behind

At the same time, the governance infrastructure lags well behind the capability infrastructure. David Shipley of Beauceron Security warned in InfoWorld that cybercriminals forced to constantly rebuild infrastructure after takedowns will find frictionless agent provisioning a huge win.

Stripe’s own fraud data reinforces the concern: one in six sign-ups across its AI services are from malicious actors, and free-trial abuse has more than doubled in the past six months.

At the same time, ‘dreaming’ introduces its own risks. If a malicious actor can influence what an agent experiences in a session, they can shape what gets consolidated into long-term memory, potentially poisoning the playbooks that govern future behavior.

The EU AI Act’s high-risk system requirements take effect in August 2026 (EU government source), three months from now, with no specific guidance issued for autonomous agents that rewrite their own operational instructions or execute financial transactions independently.

The regulatory environment remains poorly equipped to handle this shift, even as the economic incentives driving greater autonomy become overwhelming. Anthropic’s jump from roughly $9B to $30B in annualized revenue within four months illustrates how quickly enterprise demand for autonomous AI systems is growing. At the same time, the company is moving beyond simple model access toward managed environments where agents persist, coordinate, learn, and operate continuously.

The moat moved to infrastructure

That shift is also changing the competitive dynamics of the AI industry. The first phase of the AI race focused largely on model quality: which company had the best reasoning, strongest benchmarks, or fastest inference. But as capabilities converge, the strategic advantage is increasingly moving toward the infrastructure surrounding intelligence itself, including identity systems, payment rails, orchestration layers, deployment environments, and persistent memory.

In other words, the moat is moving away from the model and toward the operational ecosystem around it.

That helps explain why companies like Stripe and Cloudflare suddenly look central to the future of AI despite not building frontier models themselves. They are building the connective tissue that allows autonomous systems to participate directly in the real economy.

The internet’s first era connected information, and its second connected people and businesses. The next era increasingly appears designed around autonomous software entities capable of acting, transacting, coordinating, and learning independently. Once AI agents can remember, spend, deploy, evaluate, and improve on their own, the question stops being whether AI can assist human work and becomes how much of the surrounding operational machinery humans will continue to control directly.

The industry is no longer simply building smarter chatbots. It is building software entities with memory, identity, economic access, and operational autonomy, all of which increasingly resemble digital labor participating directly in the economy.