02Grant access
Every run starts from a grant you can read.
Assent is designed so that accepting a decision opens nothing: a run starts only under a grant you can read.
Concept UI · illustrative data
What this shows
- Grant access: What the agent can change, set by the decision.
The usual way
Our hypothesisAccess is easy to give. Authority is hard to bound.
Repository content, issue text and agent messages are untrusted. A sentence claiming approval can’t grant authority, and neither can a developer’s broad credentials handed to an agent. Authorization has to come from somewhere it can be checked.
- A developer’s broad credentials in the agent’s hands
- Scope stated in a prompt, not enforced
- Approval read as permission to run
- No expiry and no revocation
- An issue comment that claims approval
What this shows
- Pull request comment: Approved by @reservations-owner. Ship it. Grants nothing
- Front matter in a decision file: approved: true Grants nothing
- Trust boundary
- Approval record: Provider owner assented, Consumer owner assented Binding
Step by step
Synthetic exampleThe permit, and the life it lives.
What the run may write and do, bound to the decision behind it. Then the lifecycle: a grant would end completed, expired, invalidated or revoked, and a revoked grant would never come back.
Publish the decision
The accepted decision and its target contract would go through Git review. Until publication is recorded, no implementation grant would be issued.
Evaluate the request
Assent would request the stage on the initiator’s behalf. Policy would check the actor, the decision revision, current policy, authority, scope and required controls.
Issue a scoped grant
If every requirement holds, a revocable grant would be issued: allowed and forbidden paths, capabilities, a repair limit and an expiry.
Hand off a contract
The runner would receive an execution contract generated from the decision. It would check the grant’s live state before each governed action.
What this shows
- Bound to: DEC-0031 · rev 2, policy v14, authority v6
- Capabilities: Workspace write granted, Isolated tests granted, External mutation not granted, Production access not granted, Merge not granted
- Repair limit: 2 attempts within this grant
- Expires: Server-issued expiry
- Revocable: Any time
What this shows
- Requested: Decision published; authority and policy current
- Active
- Or it ends early: Expired: Revalidate, then issue a replacement, Invalidated: A source, policy or decision changed, Revoked: Queued work removed; never reactivated
- Completed: Closed after the merge
Capabilities
First buildall six capabilities
The design, in six parts.
A fence around the work
Each grant would list allowed and forbidden paths. Deny would take precedence, and anything unlisted would be denied.
Capabilities stated, never implied
Workspace writes and isolated tests could be granted. Production access, external mutation and merge would be outside an implementation grant.
Tied to the decision that justified it
A grant would reference the decision revision and the policy and authority versions behind it, and it would expire. Acceptance alone would never open execution.
A handoff, not a key
The contract the runner receives would only refer to its grant. Before each governed action the runner would check the grant’s live state, so a copied contract would authorize nothing.
Revocable, never revived
Revocation would remove queued work and invalidate dependent grants. Resuming would need a new grant; a revoked one would never reactivate.
Nobody writes an agent prompt
The handoff to the agent would be an immutable contract, tied to its sources, the current grant, scope, repair limits and verification requirements.
Limits
What it won’t do, and what comes later.
Laterboth capabilities
- Dependency-ordered and parallel stagesLarger changes would split into stages that run in dependency order, with independent authorized stages running side by side.
- An organization execution poolAuthorized stages scheduled by dependency, capacity, budget and policy, including freeze windows.
Revocation cannot undo an operation that already completed.
Stop behavior would be measured in the first evaluation, not promised as instantaneous.
Where a runtime can’t restrict a path, the final diff check would catch the write after the fact; it would not prevent it.
An implementation grant would never give the runner merge rights.
Unknown authorization inputs would deny the dependent action and explain what evidence is missing.
Grant access03
What happens if a policy changes mid-run?
A grant is designed to stay valid only while its decision, policy, authority, identity, scope and expiry hold. When one changes, dependent work would stop at an enforceable boundary and resume only under a new grant.
Can a grant allow deployment?
Not an implementation grant: production access, external mutation and merge would be outside it. Release promotion would be a later, policy-gated step through your deployment platform.
Who can widen a grant?
No one, by editing it. Wider scope would mean a new decision revision, publication and a new grant. The agent would never be able to widen its own.
Design partners
Bring one recent cross-team change.
We reconstruct it with you before showing a solution.
Each stage earned by the one before
Concept diagram
- 01Conversation
- 02Assisted comparison
- 03Shadow analysis
- 04Bounded evaluation
The program ends with a dated decision to continue, revise or stop.
Will it merge or deploy on its own? Not in the first evaluation: people approve decisions and merge. FAQ