Intent becomes safe execution here.

VerifiedX sits immediately before a real side effect. It checks the proposed action against the evidence gathered in that run, then returns a decision your agent can execute, replan from, or route upstream.

Open the quickstart Review 25-50 traces

RUN / CURRENTBOUNDARY CONTRACT

01 / INPUT Same-run evidence

Tool results, retrievals, messages, memory, upstream context.

02 / PREFLIGHT Should this run take this action?

03 / OUTPUT Decision receipt

Outcome, reasons, safe next steps, retry and routing contract.

NO TARGET-SYSTEM CREDENTIALSLOCAL / UPSTREAM RESUME

ADDITIVE RUNTIMEPRE-SIDE-EFFECTTYPED RECEIPTSFRAMEWORK AGNOSTIC

The runtime contract

A small interface. A serious control point.

Keep the agent architecture you already chose. Add VerifiedX only where a method crosses from reasoning into money movement, external communication, record mutation, or another durable side effect.

Capture

Bind the methods that create run evidence and the actions that change business state.

Preflight

Check the exact proposed action against direct, same-target facts before execution.

Resume

Continue downstream, replan locally, pass the receipt upstream, or stop the unsafe path.

Executable outcomes

Not another risk score.

A concrete runtime decision with the information the next node needs.

01 / CONTINUEAllow

The evidence positively supports the proposed effect.

02 / CONTINUEAllow with warning

The action is justified; the receipt preserves a bounded caution.

03 / KEEP WORKINGReplan required

Block this execution and return the supported next path for the same goal.

04 / STOP PATHTerminal block

Use only when the goal cannot continue safely through this route.

Bind your real seam

No orchestration rewrite.

Wrap the business methods you already own. VerifiedX observes the run and preflights only the actions you classify as consequential.

OpenAI AgentsLangGraphVercel AI SDKClaudeMCPRaw runtime

Read the integration guide

agent.tsTypeScript

import { initVerifiedX } from "@verifiedx-core/sdk";

const vx = await initVerifiedX();

vx.bindHarness(agent, {
  retrievals: {
    fetchInvoice: "invoice evidence",
  },
  actions: {
    issueRefund: "issueRefund",
  },
});

Built for operators too

Inspect the decision, not a black-box score.

Every boundary outcome can be logged with its direct reasons, safe next steps, execution status, and resume contract.

Exact reasons

See which observed fact supported or contradicted the pending action.

Safe next steps

Return an actionable alternative that keeps the original goal alive.

Execution reports

Separate what was approved from what actually executed successfully.

Replayable receipts

Give downstream agents, upstream orchestrators, and humans the same decision record.

Start at one boundary

Find the action worth protecting first.

Use 25-50 anonymized historical traces to select the seam, estimate the allow / replan split, and define a production-relevant pilot.

Email founders for a trace review No production credentials required.