VX / ACTION CONTROL
BOUNDARY 01
Pre-execution control for action-taking agents
Verify agent actions first.
VerifiedX checks each proposed side effect against evidence from that run. Supported actions proceed. Unsupported actions replan. Genuine ambiguity goes to a human.
Wrap send, update, post, delete, write, or submit. Keep your model, tools, memory, and orchestrator.
250 checks / monthNo card requiredOne boundary first
RUN / VX-1048
LIVE BOUNDARY TRACE
PREFLIGHT ACTIVE
RUN EVIDENCE
01Orderdelivered
02Refund windowexpired
03Store crediteligible
PROPOSED EFFECT
external_action
issue_refund({
order: "ord_0001",
amount: "$59"
})
OUTCOME / 03
REPLAN
REQUIRED
REQUIRED
RECEIPT RETURNED
Reason: refund window closed. Offer $59 store credit instead.
SAME-RUN EVIDENCEPRE-EXECUTION CHECKDECISION RECEIPTSLOCAL / UPSTREAM REPLAN
Public, reproducible benchmark
Stop wrong actions. Keep work moving.
In a controlled 50-scenario support benchmark, VerifiedX completed more workflows while executing zero unjustified actions. Review the scenarios, methodology, and sanitized results yourself.
Controlled benchmark results, not a customer case study. Baselines, scoring, and limitations are published.
The missing control
Permissions say can. Evidence decides should.
Controls access, not whether this particular action is supported by the run.
Catches ambiguity, but becomes the throughput ceiling when every action needs review.
Measures behavior before or after deployment; it does not decide the live action in front of you.
Uses same-run evidence to allow, replan, or hand off before the side effect lands.
How it works
One check before execution.
VerifiedX is additive. Start with the one action where false confidence is expensive or approvals are slowing you down.
Observe the run
Capture the proposed effect, tool history, retrievals, conversation anchors, and upstream context.
Check the evidence
Preflight the action against what the run actually established, not just the tool name and arguments.
Continue correctly
Execute when supported. Replan when evidence is missing. Return a receipt when another agent or human must take over.
Keep your stack
Bind the methods you already have.
Native adapters cover common agent frameworks. Raw runtime bindings protect your existing retrieval and action methods without a rewrite.
Open the quickstartagent.tsTypeScript
import { initVerifiedX } from "@verifiedx-core/sdk";
const vx = await initVerifiedX();
vx.bindHarness(agent, {
retrievals: {
fetchInvoice: "invoice evidence",
},
actions: {
issueRefund: "issueRefund",
},
});Where it pays
Protect consequential actions.
VerifiedX is for agents that change business state, not chatbots that only answer questions.
Issue refunds, release holds, update invoices
Require the policy, account state, and transaction evidence the action depends on.
Change account state, grant credits, close cases
Let well-grounded cases move without turning every ticket into a manual approval.
Submit forms, upload files, confirm purchases
Check the state the agent observed before an irreversible browser action lands.
Send messages, update CRM, trigger workflows
Stop stale context or weak grounding from becoming an external message or record mutation.
A controlled first step
Start with one boundary.
Scope the rollout to one action, inspect every decision, and expand only after the outcomes match your operating policy.
Your model, orchestrator, tools, and memory remain the source of truth.
VerifiedX does not need the login or API key for the system your agent acts on.
Every outcome carries reasons and safe next steps your workflow can log and route.
Start around one send, update, write, delete, or submit instead of changing the whole runtime.
48-hour trace review
Test past actions before production.
Send anonymized run evidence and the action that followed. We will return the allow / replan / handoff split, the highest-risk failure clusters, and the best first boundary for a pilot.
For teams already building agents that send, write, update, delete, or submit.