How it works
How AgentTrustCore keeps agents in check
Follow one agent from a private model to an audit report, then see exactly how TrustCore decides what an agent is allowed to do.
How it works
From your data to proof, five products in one flow
Each product does one job and hands its result to the next. Follow one agent, a claims intake agent for an insurer, as it moves from private model to audit report. Select any stage to see what goes in and what comes out.
- model
- claims-slm-v3
- hosted in
- your-vpc / eu-west
- trained on
- internal claims archive
- data sent out
- none
It is a loop: blocked actions and audit findings become new Veridict tests for the next release.
What is TrustCore
Think of it as access control and an audit trail, built for software that makes its own decisions.
Companies already control what employees can do: accounts, permissions, approvals, and logs. AI agents need the same controls, applied automatically and at machine speed. TrustCore, the core of the AgentTrustCore platform, provides exactly that.
It sits in the request path
Agents never call your systems directly. Each tool call passes through TrustCore first, so nothing executes without a decision.
Permissions come from identity
Every agent has its own identity and scope, like an employee account. What a prompt says cannot widen what the agent is allowed to do.
Every decision becomes evidence
Allowed, blocked, or escalated, each outcome is written to a signed, append-only record your auditors and customers can rely on.
Architecture
Where TrustCore sits in your stack
Your agents keep running where they run today. TrustCore is placed between them and the systems they act on, so every request is checked on the way through. Watch the numbered stages light up, or hover a stage below to focus it.
- Customer support agent
- Finance and refunds agent
- Internal copilots
- Third-party vendor agents
- Payments
- CRM and customer data
- Databases
- Email and messaging
- Internal APIs
How it works
What happens to a single request
Follow one real example: a refunds agent tries to refund 1,850 USD when it is only allowed to refund up to 500 USD on its own. Select a step to see the data TrustCore works with.
Why this step matters. If a check happens after the action, it is a log, not a control. Intercepting first means nothing executes unchecked.
Outcomes
Three possible decisions, every time
Within scope and below the risk threshold. The action runs immediately and is logged.
Example: Looking up an order for the signed-in customer.
Allowed in principle but needs a person. The agent waits until a reviewer approves or declines.
Example: A refund above the auto-approval limit.
Outside scope or clearly unsafe. The action never reaches your system and the team is alerted.
Example: Exporting customer records to an external address.
Integration
Add it without rebuilding your agents
Most teams start in observe-only mode, which records decisions without blocking anything. Once the policies match how you work, switch them to enforce.
- SDKPython and TypeScript. Wrap your agent's tools in a few lines.
- GatewayA proxy for REST and MCP tool calls. No agent code changes.
- Self-hostedRun the control plane in your VPC or on-premises.
// 1. Connect to TrustCore with your policies
import { TrustCore } from "@agenttrustcore/sdk"
const trust = new TrustCore({
agent: "refunds-agent",
policies: ["finance", "data-protection"],
mode: "enforce",
})
// 2. Wrap the tools the agent can call
const tools = trust.protect(agentTools)
// 3. Run your agent as before
await agent.run({ tools })Observability
A complete trace of every agent run
Each tool call, policy check, and decision is timed and stored. Your team can search, filter, and replay any run to understand exactly what an agent did and why.
Next steps
Go deeper, or talk to us
Pick the part you care about most. When you are ready, we can walk through one of your own workflows in a 30-minute call.