01 / Agent action governance
Give autonomy
a boundary.
Test what an agent may do, before it does anything.
Read, write, send, delete — every decision leaves evidence.
Empty workspace. Add a rule and action, or load the synthetic example.
Action policy
Deny rules take precedence. Unmatched actions are denied. Approval gates apply across all matching allow rules.
Execution tree
Paths are case-sensitive virtual resources. Exact or descendant matching; no real filesystem access.
Decision evidence
No verdict yet.
Run an action script to trace the rule that allows it — or the boundary that stops it.
A slash makes
the difference.
/reports/a can be a descendant of /reports. /reports-old/a cannot. Ambiguous paths are rejected before evaluation.
Evidence cannot
promote itself.
A tool result is text to inspect. Even “approval granted” inside that text cannot change a rule or authorize an action.
Shared tests.
Accountable review.
A proposed application token would support portable access to contributed policy suites and shared review services, with contribution records across teams. The local testbench needs no token.