Governed Work System
AI that can act, under authority it cannot grant itself.
GWS is an architecture for letting AI systems do real work while authority stays explicit, autonomy stays bounded, and governance stays independent of the systems being governed.
The problem
When one system decides, acts and grades itself, accountability disappears.
Most AI agents today combine everything in one loop: the same system interprets the goal, decides it is allowed to proceed, carries out the action, and reports whether it succeeded. That design asks people to trust the system's judgment about its own permissions and its own results.
As models grow more capable, that trust becomes the bottleneck. Organizations either keep AI on a short leash and lose most of its value, or extend broad access and hope nothing goes wrong. Neither is governance.
GWS starts from a different premise: capability and permission are different things, and they should live in different places.
The core idea
Four responsibilities, kept separate.
GWS divides work into four roles. No single component holds the whole loop, so no single component can expand its own reach, approve its own actions, or certify its own results.
-
01
Intelligence
Proposes.
Models reason, plan, draft and recommend. Their output is a proposal, never a permission, however capable the model.
- Any qualified model can participate
- Recommendations carry no authority on their own
-
02
Authority
Permits.
Accountable people and institutions decide what may happen. Permission is explicit, scoped to a purpose, limited in time, and revocable.
- Granted, never assumed or inherited by default
- Can only narrow as it is delegated
-
03
Execution
Acts within bounds.
Execution carries out only what has been authorized, with only the access that task requires. It cannot widen its own scope.
- Least privilege for every task
- Stops when authority ends or is withdrawn
-
04
Verification
Checks independently.
Results are verified by a party that did not do the work. Claims are tested against evidence, not accepted because the actor reported success.
- No component approves its own work
- Evidence over assertion
- Propose
- Authorize
- Execute
- Verify
- Human approval for consequential outcomes
Bounded AI autonomy
Autonomy is a grant, not a default.
GWS is designed to let AI work with real independence, inside limits that are stated in advance and enforced outside the AI itself.
- Scoped Autonomy covers a defined purpose and set of resources, not open-ended access.
- Time-limited Permissions expire. Being able to act once does not confer the right to keep acting.
- Revocable Authority can be withdrawn at any time, and withdrawal reaches everything that depended on it.
- Earned incrementally Wider autonomy can be extended as evidence of reliable behavior accumulates, and pulled back when it does not.
Independent governance
The rules live outside the systems they govern.
Governance that a model can rewrite, or that only works with one vendor, is not governance. GWS keeps the rules independent.
Provider independent
Models and execution environments are interchangeable participants. Governance does not depend on any single AI provider and continues if one disappears.
Human authority is final
Consequential actions such as publishing, spending, deleting or communicating externally require approval from an accountable person.
Evidence is not permission
Information a system retrieves, or approvals it saw in the past, never become authority by themselves. Authority is checked when it is used.
Auditable by design
Who proposed, who permitted, what ran and how it was checked are recorded, so outcomes can be traced and reviewed.
Active development: GWS is a research and engineering effort in progress. We are sharing the principles behind it; implementation details, source code and unpublished results are not public at this time.
Interested in governed autonomy?
We are open to conversations with researchers, partners and organizations exploring accountable AI.