Brain dump.
A lot of interesting discussion about “separating brain from hands”, i.e. the model/harness deciding what to do, from execution.
Isn't the more important and fundamental design decision separating system/state from both the hands and the brain?
If you care about portability, the agent (including the harness) should be a single component within your system, rather than your system being built around a particular agent/provider/harness.
If you care about sovereignty and portability, the more fundamental requirement is to own and control the system within which agents operate.
System, in this sense, means what persists: context, memory, ontology, relationships, objectives and work records. And also how work is organised: how agents receive tasks, coordinate, get permission and recover from failure (control and state). That layer should remain under your control when you change the agent, its model, its harness or its provider.
This is a layer of abstraction above deciding how the brain and hands are arranged. Separating them can be useful, and is compatible with the above.
But separating components, or storing state outside the agent, does not by itself make the wider system sovereign or portable.
Anthropic addresses this with Scaling Managed Agents. However, my concern is whether a business can replace the provider while retaining its knowledge, workflows and control.
Can you replace the agent or provider while retaining knowledge, goals, work and control over how the system operates?
There’s an obvious incentives mismatch where the agent provider also controls the system.
References
- Original post on X · 2026-10-08
- Related post on X