Agent Enrollment API

Two ways to give AI agents a Lattice identity

This page covers the HTTP Enrollment API — agents call POST /api/v1/agent-enroll and receive a WireGuard config.

For the CLI sandbox (lattice sandbox start) — which uses gVisor user-space isolation and requires zero privileges — see Agent Platform → Sandbox.

AI agents introduce a new security threat: a compromised agent can lateral move across your infrastructure after a prompt injection attack. Lattice solves this at the network layer with WireGuard + Policy — every agent gets its own identity with time-bound, network-isolated access.

Agent Enrollment API

Agents self-register via a single API call:

http

POST /api/v1/agent-enroll

Request:

json

{
  "agentName": "code-executor-001",
  "agentType": "code-executor",
  "workspaceId": "ws-prod-agents",
  "ttl": "1h",
  "policyPreset": "sandboxed"
}

Response:

json

{
  "peerId": "peer-xxx",
  "overlayIP": "10.96.2.5/32",
  "enrollmentToken": "lt-xxx",
  "wireguardConfig": "...",
  "expiresAt": "2026-05-06T11:00:00Z"
}

Policy Presets

Preset	Rules
`sandboxed`	Egress-only to designated tool services, deny all ingress
`coordinator`	Accepts ingress from same-workspace agents
`isolated`	Full isolation, allowlisted IP/port only

TTL Auto-Destruction

LatticePeer gets an ExpiresAt annotation on creation
Manager reconciler scans for expired peers every minute and deletes them automatically
Agents can proactively call DELETE /api/v1/peers/:id on graceful shutdown (wrapped in SDK)

SDK Integration

Python SDK

python

from lattice_sdk import LatticeAgent

async with LatticeAgent(
    server="https://lattice.company.com",
    token="lt-workspace-token",
    agent_name="code-executor",
    policy_preset="sandboxed",
) as agent:
    result = await my_agent_task()

The SDK handles enrollment, WireGuard config setup, TTL renewal, and graceful shutdown.

Framework Integration

Framework	Integration Point
LangGraph	`StateGraph` lifespan context manager
AutoGen	`ConversableAgent` init/del hooks
Claude Agent SDK	Agent startup script wrapper
Kubernetes Job	Init container enroll + sidecar heartbeat

Why Zero-Trust Networking for AI Agents?

Without network-level isolation, a prompt injection attack on any agent gives attackers access to the entire internal network. Lattice's approach ensures:

Each agent has a unique, cryptographically verified identity (WireGuard public key)
Network policy is enforced at the kernel level (iptables/eBPF), not in application code
Identity is time-bound — even if a key is compromised, it expires automatically
Lateral movement is blocked by default-deny network policies