You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: README.md
+12-1Lines changed: 12 additions & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -23,6 +23,17 @@ This closes the confused-deputy gap where an agent can misuse delegated credenti
23
23
-**Deterministic binding**: authority is tied to runtime evidence, not only identity.
24
24
-**Provable controls**: each decision can emit signed proof events for audit pipelines.
25
25
26
+
### Why not just use IdP directly?
27
+
28
+
You should still use Entra/Okta/OIDC for identity and token issuance. `predicate-authority` adds the runtime control layer those systems do not provide by default for AI agents:
29
+
30
+
- pre-execution allow/deny checks right before each sensitive action,
31
+
- binding authority to current `state_hash` and `intent_hash`,
32
+
- optional required verification labels from runtime checks (currently web-agent only via [predicate-sdk](https://github.com/PredicateSystems/sdk-python) integration),
33
+
- fail-closed local enforcement and per-decision proof events.
34
+
35
+
In practice: IdP answers **who the principal is**, while `predicate-authority` answers **whether this exact action is allowed right now in this state**.
36
+
26
37
## Repository Components
27
38
28
39
| Package | Purpose |
@@ -38,7 +49,7 @@ Implemented in this repository:
38
49
- local pre-execution `ActionGuard.authorize(...)` and `enforce(...)`,
39
50
- signed local mandates with TTL (`LocalMandateSigner`),
40
51
- policy evaluation with deny precedence and required verification labels,
0 commit comments