AI Agent Overview

AI agents represent a fundamentally new class of non-human identity. Unlike traditional service accounts that execute deterministic code paths, AI agents make autonomous decisions, interact with unpredictable external inputs, and can incur unbounded costs if left ungoverned. VeraID provides purpose-built controls for managing the unique risks that AI agents introduce to your organization.

Why AI Agents Need Special Identity Management

Traditional NHI governance was designed for predictable, deterministic systems. AI agents break every assumption those systems were built on.

Autonomous Decision-Making

AI agents interpret prompts, select tools, and take actions without explicit human instruction for each step. A single prompt can trigger a chain of API calls, database queries, and external service interactions — all decided by the model at runtime. Without governance, there is no way to predict or constrain what an agent will do.

Unpredictable Costs

Every LLM call consumes tokens, and token consumption translates directly to spend. An agent caught in a reasoning loop, processing unexpectedly large inputs, or making recursive tool calls can exhaust budgets in minutes. Traditional rate limiting is insufficient because a single “request” to an agent can fan out into dozens of LLM calls internally.

Prompt Injection Risks

AI agents process untrusted input — user messages, retrieved documents, tool outputs — and any of these can contain adversarial instructions designed to override the agent’s intended behavior. A successful prompt injection can cause an agent to exfiltrate data, bypass access controls, or execute unauthorized actions, all while appearing to function normally.

Lack of Visibility

Most organizations have no centralized view of which AI agents are running, what models they use, what data they access, or how much they spend. Agents are often deployed by individual teams with ad-hoc credentials and no audit trail, creating blind spots that grow as adoption accelerates.

Danger

An ungoverned AI agent with broad credentials is a security incident waiting to happen. It combines the access of a service account, the unpredictability of a human user, and the speed of automation — with none of the oversight controls that apply to any of those categories.

---

Agent Identity Model

VeraID models AI agents as first-class identities with type AI_AGENT. Every agent identity carries a standard set of NHI fields (name, status, tags, risk score) plus an agentConfig object that encodes agent-specific governance controls.

Agent Configuration Schema

{
  "id": "idt_9c3a8b7e-4f21-4d6a-b8e1-a2c5d9f07e3b",
  "name": "customer-support-agent",
  "type": "AI_AGENT",
  "status": "ACTIVE",
  "agentConfig": {
    "provider": "openai",
    "modelIdentifier": "gpt-4o",
    "capabilities": [
      "chat-completion",
      "function-calling",
      "retrieval"
    ],
    "dataAccess": {
      "allowed": [
        "customer-tickets",
        "knowledge-base",
        "product-catalog"
      ],
      "denied": [
        "financial-records",
        "employee-pii",
        "source-code"
      ]
    },
    "monitoring": {
      "logAllRequests": true,
      "logResponses": true,
      "requireApproval": false
    },
    "tokenBudget": {
      "daily": 50.00,
      "monthly": 1000.00,
      "perRequest": 2.00
    }
  }
}

Configuration Fields

Field	Type	Description
provider	string	LLM provider: openai, anthropic, azure_openai, or custom
modelIdentifier	string	Specific model used by this agent (e.g., gpt-4o, claude-sonnet-4)
capabilities	string[]	Declared capabilities: chat-completion, function-calling, retrieval, code-execution, image-generation
dataAccess.allowed	string[]	Data sources this agent is permitted to access
dataAccess.denied	string[]	Data sources explicitly denied, regardless of other policies
monitoring.logAllRequests	boolean	Log every prompt sent to the LLM
monitoring.logResponses	boolean	Log every response returned from the LLM
monitoring.requireApproval	boolean	Require human approval before executing sensitive operations
tokenBudget.daily	number	Maximum spend per day in USD
tokenBudget.monthly	number	Maximum spend per month in USD
tokenBudget.perRequest	number	Maximum spend per individual request in USD

Tip

Set dataAccess.denied explicitly for sensitive data categories even if they are not in the allowed list. Deny rules act as a safety net — they block access even if a misconfigured policy would otherwise grant it.

---

Supported Agent Frameworks

VeraID integrates with the major AI agent frameworks and LLM providers. The integration method depends on the framework’s architecture.

SDK Callback Integration

These frameworks support native callback hooks that VeraID intercepts for policy evaluation, budget enforcement, and audit logging.

Framework	Integration Method	Key Features
LangChain	SDK callbacks	Intercept chain execution, tool calls, and LLM invocations via LangChain callback handlers
LlamaIndex	SDK callbacks	Monitor query pipelines, retrieval steps, and response synthesis through event hooks
Semantic Kernel	SDK callbacks	Govern function calls, planner execution, and connector usage via Semantic Kernel filters

import { VeraIDClient } from '@veraid/sdk';
import { VeraIDLangChainCallback } from '@veraid/sdk/langchain';

const client = new VeraIDClient({
  apiKey: process.env.VERAID_API_KEY,
  agentId: process.env.VERAID_AGENT_ID,
});

// Attach VeraID as a LangChain callback handler
const agent = new AgentExecutor({
  agent: myAgent,
  tools: myTools,
  callbacks: [new VeraIDLangChainCallback(client)],
});

API Proxy + SDK Integration

These frameworks are governed through a combination of API-level proxying (for LLM calls) and SDK-level instrumentation (for tool and task orchestration).

Framework	Integration Method	Key Features
AutoGPT	API proxy + SDK	Route LLM calls through VeraID gateway; SDK monitors task planning and execution
BabyAGI	API proxy + SDK	Proxy model calls for budget and injection checks; SDK tracks task queue and results
CrewAI	API proxy + SDK	Gateway intercepts all crew member LLM calls; SDK governs inter-agent communication

Credential Injection

For direct LLM API usage, VeraID manages credentials and injects them at runtime. No code changes are required beyond pointing to the VeraID credential endpoint.

Provider	Integration Method	Key Features
OpenAI	Credential injection	VeraID vaults and rotates OpenAI API keys; SDK injects credentials per-request
Anthropic	Credential injection	Managed Anthropic API keys with per-agent budget isolation
Azure OpenAI	Credential injection	Azure AD token management with VeraID policy enforcement

Workflow Platform Integration

Low-code and automation platforms connect via webhooks and secrets synchronization.

Platform	Integration Method	Key Features
n8n	Webhook + secrets sync	VeraID syncs credentials to n8n credential store; webhooks trigger policy evaluation
Zapier	Webhook + secrets sync	Managed Zapier connection credentials with usage tracking and auto-rotation

Custom Agents

Any agent or LLM-powered application can integrate with VeraID through the REST API and SDK.

Approach	Integration Method	Key Features
Custom agent	REST API + SDK	Full API access for budget checks, prompt scanning, credential retrieval, and audit logging

import { VeraIDClient } from '@veraid/sdk';

const client = new VeraIDClient({
  apiKey: process.env.VERAID_API_KEY,
  agentId: process.env.VERAID_AGENT_ID,
});

// Before each LLM call
await client.agents.checkBudget(estimatedCost);
await client.agents.guardPrompt(userInput);

// After each LLM call
await client.agents.recordSpend({
  amount: actualCost,
  model: 'gpt-4o',
  operation: 'chat-completion',
});

Note

All integration methods produce the same audit trail format. Regardless of whether you use SDK callbacks, the API proxy, or direct REST calls, every agent action is logged with the identity, timestamp, operation, cost, and policy evaluation result.

---

What’s Next

• Agent Gateway — Understand how VeraID intercepts and governs LLM API calls
• Budget Controls — Configure spend limits and cost tracking for AI agents
• Prompt Injection Detection — Protect agents from adversarial prompt manipulation
• Approval Workflows — Require human review for sensitive agent operations
• MCP Server Governance — Manage Model Context Protocol servers and tool access