Review the exact payload
Show old and new values, recipients, artifacts, affected records, risk, and policy reason—not just a tool name.
Onterix Actions
From agent intent to governed action.
Reads create confidentiality risk. Writes add integrity, financial, operational, and reputational risk. Onterix is designed to make every mutation explicit before it becomes real.
GOVERNED TRANSACTION
Interactive trace“Prepare the Acme renewal brief, update the forecast, and send the customer follow-up.”
STEP 01
Bind the real actor and client
Maya asks Codex to prepare a renewal brief and update Salesforce. Onterix binds her identity, tenant, client, and exact capability versions before any source is touched.
Identity established
- Actor
- Maya Chen · Revenue Ops
- Client
- Codex · verified
- Capabilities
- 4 requested · 3 visible
STEP 02
Inspect the data boundary
Source labels and configured protection providers classify the assembled context. Customer PII is tokenized and restricted pricing is kept on the approved model route.
Permitted derivative created
- Sources
- Drive · Slack · Salesforce
- Findings
- 2 PII · 1 confidential
- Decision
- Transform + force route
STEP 03
Approve the exact mutation
The proposed CRM change and customer email are separate transactions. Policy allows the forecast update and requires a revenue-owner approval for the external message.
Payload sealed for review
- CRM diff
- Stage 2 → Stage 3
- 1 recipient · 1 attachment
- Approval
- Revenue owner · 30 min
STEP 04
Revalidate, execute, and prove it
Immediately before execution, Onterix rechecks policy, permissions, source revisions, and the approved payload. The result is idempotent and the full decision chain becomes evidence.
Executed with evidence
- Execution
- 2 actions · idempotent
- Policy
- Revalidated at 14:32:08
- Trace
- otr_01JQ9F…
Exact-action governance
Approve one mutation—not a blank check.
An approval is bound to the exact capability, payload, destination, source revisions, policy, and protection versions. Material change supersedes the approval.
01Normalize the proposed mutation and recipientsintent
02Resolve actor, agent, client, source, and destinationcontext
03Inspect data and classify action riskprotect
04Compute a human-readable payload and diffpreview
05Evaluate policy and acquire eligible approvalapprove
06Seal versions, hashes, and source revisionsbind
07Revalidate immediately before executionrecheck
08Execute idempotently and record evidenceprove
Transaction semantics
The controls serious actions require.
A confirmation dialog is not an approval system. Enterprise actions need durable state, separation of duties, revision safety, and proof of what actually happened.
Agents cannot approve themselves
Eligible reviewers are resolved through tenant roles and policy. Approval never expands what the actor could otherwise do.
Stale approval is not permission
Changed payload, destination, source revision, policy, ruleset, or transformation invalidates the prior decision.
Retry without duplicate impact
Execution keys and durable state distinguish a safe retry from a second mutation, message, task, or charge.
One request, multiple transactions
Govern each consequence independently.
“Update the opportunity and email the customer” is not one action. Onterix separates the CRM mutation from external delivery so policy can allow one, transform another, and require approval for the highest-risk consequence.
Forecast update
Permitted fields, current record revision, old/new values, owner authority, and idempotent mutation.
Customer follow-up
Recipients, body, attachments, inherited classification, protection decision, and separate reviewer.
Connected evidence
One parent intent with two distinct decisions, approvals, execution states, and outcomes.
Design partner workflow
Choose one write that matters. Govern it end to end.
The strongest private-preview pilots begin with a concrete mutation that is valuable enough to automate and sensitive enough to require real controls.