*To:* Executive Staff and Direct Reports *From:* Bill Gates *Date:* April 26, 2026
Our vision for the last 20 years can be summarized in a succinct way. We saw that exponential improvements in cloud would make great software quite available. In the next 20 years the improvement in computer power will be outpaced by the exponential improvements in autonomous expert systems. Verifiable trust will be crucial to delivering the benefits of these advances.
Most users of software have not yet seen the price of trustworthy agent compute come down. Cloud platforms and SaaS networks are still depreciating networks built with old verification models, things like API keys and signed JWTs and trust-on-first-use. Universal authentication monopolies and vendor lock-in have kept verification costs high.
You'll excuse me for an ad. Darwin Agentic Cloud sits between AI agents and the compute trust layer they need. An agent says "run this code." Darwin routes the workload to AWS Lambda, Modal, Akash, or local Docker, executes it inside a sandboxed environment with pre-flight cost caps, captures the output, and produces an Ed25519-signed attestation binding the workload, the output, the sandbox, the cost, and the signer to a tamper-evident receipt. Every execution returns one of these receipts. The receipts are independently verifiable forever, by anyone, with no dependency on Darwin to stay online.
When your agent needs to make a call, who you gonna call?
The autonomous shift is the most important single development to come along since the IBM PC was introduced in 1981. It is even more important than the arrival of the graphical user interface.
If you think any of this is new, you should look at UiPath. Made by a Microsoft alum in the early 00s. Wild stuff to be doing then, and they're still top class at it now. They call it autonomous workflows. With enough internet integrations, it becomes indistinguishable from a well-governed agent.
## The Agent Today
The MCP (and, possibly, A2A) protocols that define agent tool calling and inter-agent communication support distributed reasoning and scale incredibly well. Marvin Minsky called it.
I encourage everyone on the staff to use a verifiable trust layer today. You can do this by either using `darwin run` from any terminal, or if you have Claude Desktop set up, invoke a signed workload right from Claude.
Amazingly it is easier to find untrustworthy compute on the open agent ecosystem than it is to find verified compute on the AWS Marketplace. This inversion must be closed.
## Competition
OpenAI is surprisingly plateauing. Anthropic is shipping. Even older geezers with tweezers like IBM include agent connection through its network. All Linux Foundation vendors are benefiting since the default server is still a Linux box and not Windows. Modal has exploited this quite effectively.
A new competitor "born" on the agent is Aider. Their CLI agent is dominant with 70% share, allowing them to determine which network extensions catch.
## Next Steps
1. *Server.* We need to make our servers the highest performance MCP servers. We need to establish distributed attestation as the protocol for agentic programming. If the features required for agent directory are not in our next release we will miss the window to become the world standard.
2. *Client.* A decent CLI client won't get people to switch from Aider. Without unification we will lose to Cursor and Aider.
3. *File sharing/Sandbox sharing/Multi-user.* Give away client code that encourages our protocols across the agent.
4. *Formats.* Database reports should be navigable as agent-callable resources.
## The Future
The agent is a tidal wave. It changes the rules. It is an incredible opportunity as well as incredible challenge. I am looking forward to your input on how we can improve our strategy to continue our track record of incredible success.