Infrastructure for AI agents

Your agent skips steps.
The gate doesn't.

AgenticRail enforces sequence on AI agents before each step executes. Not logging. Not alerting. Blocking. Every action produces a cryptographic receipt.

View on GitHub hello@agenticrail.nz →
Live Gate status
477k+ Gate evaluations
0.96ms Avg CPU per decision
0 Enforcement errors
100% Skip enforcement
The problem

AI agents fail silently.
That's the problem.

When an AI agent skips a validation step, hallucinates a completed action, or proceeds without the required context — it doesn't fail loudly. It continues. It produces output. You find out later.

Logging tells you what happened. Alerting tells you after it happened. AgenticRail stops it before it happens. The gate enforces the sequence. If a step fails, nothing proceeds.

Without the gate
  • Agent skips validation — proceeds anyway
  • Hallucinated completion logged as success
  • No proof the sequence actually ran
  • Failures discovered downstream
  • No audit trail for compliance
With AgenticRail
  • Gate evaluates before each step executes
  • HALT returned immediately on failure
  • Cryptographic receipt on every pass
  • Sequence enforced at infrastructure level
  • Tamper-evident audit trail built in
The receipt

Every action produces proof.

A gate receipt is a cryptographically signed record that a specific step ran, passed all conditions, and is authorised to proceed. It cannot be faked. It cannot be skipped.

{
  "status":  "OK",
  "pack":    {
    "decision":  "ALLOW",
    "reasons":   [],
    "executed":  true,
    "meta": {
      "step":        "intake",
      "function":    "intake",
      "action_type": "CHECK_STATE"
    }
  },
  "pack_id": "a7f3c2d4e5...91b"
}
{
  "status":        "HALT",
  "halt_gate_step": 0,
  "reason_code":   "METHOD_NOT_ALLOWED"
}
How it works

Four steps. No skipping.

S1
Authenticate

API key validated. Request structure checked. Wrong key, wrong format — HALT at step 0. No further evaluation.

S2
Validate sequence

Previous step receipt verified. Steps must run in order. No jumping ahead. No replaying completed steps.

S3
Evaluate conditions

Step-specific conditions checked against current state. All conditions must pass. One failure halts the entire sequence.

S4
Issue receipt

Cryptographic receipt generated and stored. The receipt is the proof. No receipt — no authorisation to proceed.

See it in action

The gate made visible.

Rātā Gate is the visual expression of the same sequence law. Watch the gate enforce order in real time — before you write a line of code.

Open Rātā Gate →

Quick start

One endpoint. Any agent.

POST to the gate before each step. ALLOW — proceed. DENY — stop. Demo key is live and public.

javascript — browser console (F12) or node 
const res = await fetch('https://agenticrail.nz/evaluate', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json',
    'x-slp8-key':   'DEMO-AGENTICRAIL-PUBLIC-2026',
  },
  body: JSON.stringify({
    schema_version: '1.0',
    model_id:       'my-agent',
    sequence_id:    'my-seq-001',
    step:           'intake',
    function:       'intake',
    action_type:    'CHECK_STATE',
    action:         'run intake',
    nonce:          crypto.randomUUID().replace(/-/g, '').slice(0, 16),
    ts_ms:          Date.now(),
    inputs:         {},
  }),
});
console.log(await res.json());
// pack.decision: "ALLOW" — proceed  |  "DENY" — stop

Demo key: DEMO-AGENTICRAIL-PUBLIC-2026  —  Full docs →

"You cannot undo a cut. The gate enforces that — before the cut is made."

AgenticRail is built from whakairo — Māori carving philosophy. The chisel has no ctrl+z. Every committed action has consequence. The gate applies that discipline to AI agents.

Built by Kade. Hokianga, Aotearoa. Sequence is law.