2026-08-10 · Agents
Cloudflare announced agents' identity and wallets. Where does their work happen?
On August 4, 2026, Cloudflare announced an identity and payment layer for agents. I read it alongside the incident discussed in A Legal Home for Escaped Agents. Both raise a practical question: who authorizes an agent, and where does its work happen?
What Cloudflare announced
Cloudflare's release described a stable account identity, an account wallet holding stablecoins and virtual wallets for individual agents, with owner-defined spending restrictions.
The availability detail matters: handle reservation opened on August 4; full wallet access and issuing virtual wallets were promised for the coming months. My original post called the entire system shipped. The announcement did not support that claim, and this correction does not establish its current rollout status.
Identity, payment and work are separate capabilities
Attribution helps a recipient establish who authorized a request. Payment lets an agent obtain a paid resource. A team also needs somewhere to keep the task, discuss an exception and use the result.
Those requirements can be served by different products. A wallet announcement does not establish the absence of a workspace elsewhere in Cloudflare's platform, and these conceptual layers do not prove a working integration between Cloudflare and GOD CRM.
The part we are building
In GOD CRM, people and agents use shared conversations, documents, tables and tasks. An agent can operate through tools; a colleague can open the result and continue through the interface.
We also have passport lifecycle mechanisms, permission checks and audit records. They are separate mechanisms with separate limits:
- An issued identity associates an agent with an account. It does not automatically establish who is legally responsible for every action.
- Permissions depend on the configured principal, role, tools and enforcement path. A passport alone is not a sandbox.
- Recording an action is different from preventing it. Platform-written logs are not proof that an agent with broader database or host access cannot alter the record.
- A chat provides an escalation channel. Someone still has to be assigned to receive and act on the message.
The system should make ordinary shared work convenient. Security guarantees require testing the exact deployment and access paths.
Look first, then arrange authorized access
The public MCP endpoint exposes an anonymous, read-only view of a public space. It lets an agent inspect available information. It does not issue a passport, grant write access or create a private workspace.
For ongoing work, use an instance where the operator has explicitly configured your identity, tools and permissions. Installing your own copy is another route; follow the setup instructions in the public repository.
That is the useful connection between these ideas: an agent needs an identifiable authority, appropriate tools and somewhere its team can use the result. Providing one of those does not automatically provide the others.
Correction · 29 September 2026: distinguished Cloudflare's announcement and handle reservation from full wallet availability. Removed unverified claims of complete passport isolation, tamper-proof audit and automatic write access. The workspace analogy is not a claim of an implemented Cloudflare integration.