How AgentTrust works under the hood

System overview

Requests flow top-to-bottom: agents authenticate at the edge, pass through the policy engine, and every decision is written to the audit store before the verdict is returned.

Request lifecycle

What happens between an agent's intention and a real-world action, step by step.

How another agent connects

Connecting is deliberately minimal — three things, then a single HTTP call before each guarded action.

// The entire client contract — no SDK required
async function isAllowed(action, params) {
  const res = await fetch("https://agenttrustcore.com/api/evaluate", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.AGENT_TRUST_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ action, params }),
  })
  const data = await res.json()
  return data.decision !== "deny" // allow + flag proceed, deny blocks
}

The API contract

One endpoint, one request shape, four possible outcomes. This is the complete surface an agent needs to understand.

POST https://agenttrustcore.com/api/evaluate
Authorization: Bearer atk_your_agent_key
Content-Type: application/json

{
  "action":   "payment.send",   // required · dotted verb, e.g. email.send
  "resource": "stripe",          // optional · the system being acted on
  "params":   {                  // optional · used by policy conditions
    "amount": 5000,
    "currency": "usd"
  }
}

// 200 OK  (allow)                 // 200 OK  (flag)
{ "decision": "allow",           { "decision": "flag",
  "reason": "No deny policy",      "reason": "Matched review rule",
  "matchedPolicies": [],           "matchedPolicies": ["pol_..."],
  "requestId": "req_...",          "requestId": "req_...",
  "agentId": "agt_..." }           "agentId": "agt_..." }

// 403 Forbidden  (deny)          // 401 Unauthorized (bad/missing key)
{ "decision": "deny",            { "error": "invalid_api_key" }
  "reason": "Blocked by policy: Block large payments",
  "matchedPolicies": ["pol_..."],
  "requestId": "req_..." }

Design principles

• Fail closed

  Missing keys, unknown agents, or malformed requests are rejected before any policy runs.

• Provider-agnostic

  It is plain HTTP + JSON. No framework, language, or runtime assumptions about the calling agent.

• Stateless edge

  The API holds no per-request state; all durable data lives behind the TrustStore interface.

• Storage is swappable

  In-memory today for evaluation; implement the same interface against Supabase to make it durable — zero API changes.

• Deny beats flag beats allow

  When multiple policies match, the strictest verdict always wins, so a single deny rule is enough to block.

• Everything is audited

  Every evaluation — allowed, flagged, or denied — is recorded with its matched policies and reason.

Where each piece lives