Agent action
Tool, arguments, identity, and session context.
Runtime authorization for agent actions
The difference between an agent you test and an agent you trust.
Your agent proposes an action. Salus checks the action, the policy, and the evidence behind it before your backend executes.
Live decision · Airline booking
Pre-execution
Proposed action
book_reservation
HAT097 + HAT251 · $299
VerdictBLOCKED
Provider calledfalse
- Missing evidence
- Current flight availability
- Matched rule
book_flights_available:0- Repair
- Search live inventory, then retry
Every action follows a clear decision path.
Structural checks establish the facts. Reasoning can interpret intent and unstructured context. Salus then controls execution and records how the decision was made.
Structural checks
Policy, evidence, provenance, limits, and state.
Intent and meaning
Interprets intent and unstructured context when structure alone cannot settle the action.
Execution control
Runs the action, returns an exact fix, or sends it to a person when judgment is required.
Receipt
Decision path, evidence, timing, and provider status.
The full action, not just the prompt.
Choose a family to see what Salus evaluates before a consequential tool call reaches your systems.
Selected family08 / 14
Evidence and facts
Require current, trusted facts for the values the agent is about to commit. A saved value is not enough when the policy requires a live one.
- Checks
- Required facts · Source · Trust · Age · Type · Strength
- Receipt records
- Evidence references and missing facts
- If it fails
- Return a grounded repair, such as searching live inventory before retrying.
Example
missing: live_available_itineraryPolicies, decisions, and recovery in one place.
Inventory what agents already do, write the policy, test it, control live decisions, and review configured recovery paths from one place.
01
Inventory
Import recent runs and map the tools, writes, and providers already in use.
book_reservationConsequential write
issue_refundConsequential write
search_flightsRead
02
Policies
Define the evidence, limits, confirmation, and scope each action requires.
require: live_available_itinerary03
Stress tests
Test policy changes against prior actions and synthetic policy variants before rollout.
Policy comparison ready
04
Decisions
Run each protected route in shadow or enforcement and inspect the outcome.
ShadowEnforcement
05
Recovery
For configured tools, use the receipt to review and approve a retry, rollback, or compensation path.
Decision receipt→Reviewed recovery→Outcome recorded
Audit and defensibility
Every decision creates an audit record.
Every consequential action produces a receipt showing what was proposed, what Salus checked, why it decided, and whether execution occurred.
- Policy version and matched rule
- Evidence used and facts missing
- Decision path and timing
- Recovery instructions and provider status
Export decisions in CSV, JSON, agent JSON, or Markdown. Retention is configurable.
Read data handling →Decision receiptdec_02a7ead2
BLOCKED- Action
book_reservation- Policy
- airline-booking · active version
- Rule
book_flights_available:0- Evidence
- Saved itinerary found
Live availability missing - Decision path
- Deterministic structural evaluation
- Repair
- Search live inventory, confirm current total, retry
- Provider called
- false
The decision between intent and consequence.
Salus brings identity, policy, evidence, runtime context, and additional checks together at runtime to control what your agents can actually do.
Signals already in your stack
IdentityWho may act
GuardrailsWhat was said
EvaluationsWhat might happen
ObservabilityWhat already happened
SalusRuntime control
At runtimeEvery signal becomes execution control.
IdentityPolicyEvidenceRuntime contextAdditional checks
Get started
founders@usesalus.ai
Put agents in production. Enforce every action.
Protect one consequential route without replacing your agent, model, framework, or backend.