Skip to content
Verdict

How it works

A decision between intent and execution

One call wraps the place your agent already executes tools. Verdict returns ALLOW, DENY, or REQUIRE_CHANGES before a high-risk action runs.

The integrated stack

One integration point, one consistent record from intent to evidence: the gateway that decides, the obligations that run, the chain that records, the pack a reviewer signs.

01 / Your agent

Proposes a tool call

Any framework, any model. One call wraps your executor.

OpenAI · Anthropic · LangChain · custom

02 / The runtime

Policy gateway

Deterministic Rego, evaluated in-process.

versioned bundles · shadow or enforce · ~5 ms measured

03 / The decision

Allow, deny, or send it back

ALLOWDENYREQUIRE_CHANGES

Real rules are yes-if. REQUIRE_CHANGES names what is missing.

rate quote → ALLOW · below-floor discount → DENY · $500 refund → repair, then retry

when policy requires it

04 / The obligations

The step the rule required

Consent, disclosure, approval, identity. The action retries once they return.

requireConsent · addDisclosure · addApproval · requireIdentityVerification

05 / The record

Cryptographic audit chain

Signed into a hash chain, pinned to the policy version that made it.

SHA-256 chain · Ed25519 signed · policy version pinned

06 / The evidence

Evidence pack

One file a reviewer reads end to end, chain verified inside.

JSON · one-page HTML · chain verified inside the file

Rollout

One call, then shadow, then enforce

Everything above the enforcement point (models, orchestration) and below it (payments, cases, data) stays yours.

  1. Step 1

    an afternoon

    Wrap tool execution

    One call wraps the place your agent already executes tools. No new framework, no replatform: TypeScript and React SDKs, plain HTTP from any language.

  2. Step 2

    a week or two of traffic

    Shadow on live traffic

    Verdict logs what it would have allowed, blocked, or repaired on real volume. Long enough that your risk owners are reading real decisions rather than a sample, and no customer is affected either way.

  3. Step 3

    when risk signs off

    Enforce, then evolve

    This step is a decision, not a build, and it is the one we cannot schedule for you. After it, rule changes are policy bundle updates that take effect without redeploying the agent.

The whole integration

import { VerdictClient, PolicyEnforcer } from "@verdict/sdk"; const client = new VerdictClient({ gatewayUrl });const verdict = new PolicyEnforcer(client); const result = await verdict.execute({  tool: "issue_refund",  args: { amount: 500, customer: "C-4471" },  context,            // who is acting, for whom  executeFn: runTool, // your existing executor  shadow: true,       // observe first});// ALLOW | DENY | REQUIRE_CHANGES + repairs

TypeScript shown. Plain HTTP underneath, so any language works.

Where it attaches

Pick whichever seam you already have

Verdict is agent-agnostic and model-agnostic. It cares about the action, not about which framework proposed it, and not about whether you have agents running yet.

TypeScript and React SDKs
Typed clients over a documented OpenAPI contract, with published JSON Schemas for every request and response.
MCP gateway
Sit in front of MCP servers and govern tool calls in both directions: the request going out and the results coming back.
Salesforce Agentforce
Governed through Flow and an External Services callout, so an Agentforce action is checked on the same path as everything else.
LiteLLM
A pre-call guardrail plugin, for estates that already route their model traffic through LiteLLM.
Portkey
Webhook integration for teams standardised on Portkey's gateway.
n8n
An estate poller that brings existing n8n automations under the same policy and evidence plane.
Any HTTP client
The evaluate endpoint is a documented OpenAPI contract with published JSON Schemas. No framework adoption required.

And wherever you are starting

You already run agents

Add Verdict at the one place your agent executes tools. Keep your models, orchestration and systems exactly as they are.

You are starting from zero

We co-build your first workflow on Verdict, disputes or fraud or onboarding or servicing, then your team runs it day to day.

Bring the seam and the engineer who owns it

Thirty minutes, live, on one workflow. We map where enforcement attaches in your stack and what your risk owners get out the other side.

Book 30 min

For architects and risk owners: where it runs and what leaves your environment, or the benchmark method and raw runs.

Book 30 min