Agentic Event Delivery Guardrails

Stop your AI agents from taking expensive, irreversible actions

AgentDelivery is for your business if costs are getting over the edge because of agents deciding on a harmful action. Put budgets, human approvals and instant rollback between the trigger and the action. The agents you already run won't change. The way they act will.

Budget & rate ceilings — kill runaway loops and cost spikes before they run
Human-in-the-loop approvals — pause any sensitive action until a person says yes
Idempotent delivery — a retry never double-fires (no duplicate refunds or double sends)
Automatic rollback & immutable audit — undo a bad action; see everything every agent did

Compare plans → Get Started

Developer Tools → Start free

Guardrails for every agent action — proportional, not all-or-nothing

Gartner expects 40% of enterprises to pull their AI agents by 2027 over governance gaps found only after the incident. AgentDelivery is the runtime layer that closes that gap: per-agent budgets, human approval on sensitive actions, atomic rollback, and a full audit trail — enforced before the action runs, calibrated to each agent’s autonomy.

How it works

Four controls, enforced before your agent acts

Every event runs the same path — authenticated, checked against your policy, executed once, and recorded. Your agent can’t act outside it.

01 · Trigger

Authenticate the trigger

Only events from sources you trust get in. Each agent acts under its own identity and scope — not a shared key.

02 · Gate

Gate the action — your circuit breakers

Budget, rate, and content gates evaluated before anything runs. Spend ceilings and loop limits stop a runaway agent before it fires.

03 · Execute

Execute once, or roll back

Idempotent delivery means a retry never double-fires. Multi-step actions run atomically — if one step fails, the rest reverse.

04 · Audit

Prove what happened

Every decision, approval, and action recorded immutably — the audit trail you’ll want before the incident review, not after.

Agentic Event Delivery Guardrails

Control What Agents Can Do, Before They Do It

Per-agent credentials. Policy enforcement before execution. Atomic bundles with rollback. OTEL traces for every action. Give AI agents and automated systems access to webhook-driven actions with full governance and auditability.

Included on all plans

See Plans →

Agent SDKs

Typed execute endpoint

Agent SDKs provide a typed execute endpoint with per-agent credentials. Each agent operates under its own identity and policy scope.

Policy Gates

Budget, rate & content rules

Enforce spend limits, throughput caps and payload validation before any agent action executes. Non-compliant requests are blocked with a clear reason.

Bundles

Atomic multi-action execution

Group multiple agent actions into a single bundle. If any step fails, preceding steps are rolled back automatically.

Approval & Traces

Human-in-the-loop & OTEL

Require human approval before sensitive actions proceed. Every execution is recorded as an OTEL trace for full auditability.

Run the real guardrails free, in production

Watch budgets, approvals, and rollback fire on real events — before you trust an agent to act at scale. No simulation, no credit card.

1,000 events / month
1 endpoint, full delivery lifecycle
The real Agent Control Plane, capped

Start free → See plan limits →

Get Started

Connected in minutes

Point your source at AgentDelivery, set a policy, and govern every event before it reaches your application.

Route webhooks through AgentDelivery

Replace your endpoint URL with an AgentDelivery ingest URL. Works with Stripe, GitHub, Shopify, or any HTTP source.

# Stripe Dashboard → Webhooks → Add endpoint
Endpoint URL: https://agentdelivery.io/v1/webhooks/stripe/{source-id}

# Or register a source via API
curl https://agentdelivery.io/v1/sources \
  -H "Authorization: Bearer dr_live_..." \
  -d '{"name":"stripe","type":"stripe"}'

Define a policy gate

Set budget limits, rate caps, and required fields. Non-conforming events are blocked before they reach your app.

{
  "budget_eur_day": 500,
  "rate_per_minute": 100,
  "require_fields": ["customer_id", "amount"],
  "block_on_violation": true
}

Execute with the Agent SDK

Act on governed events from your app or an AI agent. Every call is idempotent, policy-gated, and OTEL-traced.

from duerelay.agent import Client

client = Client(agent_key="dar_live_...")
result = client.execute("process-payment", {
    "order_id": event["data"]["id"],
    "amount": event["data"]["amount"],
})
# idempotent · policy-gated · OTEL-traced

Read the Quickstart → Browse Docs →

Developer Tools

See and fix exactly what your agents delivered — every payload, retry, and endpoint, in one place.

Webhook Inspector

Inspect the payload and every delivery attempt behind each agent event.

View tool →

Replay Failed Webhooks

Re-fire any failed delivery instantly to validate a fix — idempotent, no duplicate side effects.

View tool →

Delivery Timeline

Track every stage from agent event to confirmed delivery.

View tool →

Endpoint Health Monitor

Spot failing endpoints and delivery errors before your agents do.

View tool →

Govern what your agents do — not just what they can reach

Access controls, such as IAM, OAuth scopes, API keys decide what an agent can reach. They say nothing about what it’s allowed to do once it gets there.
AgentDelivery governs the action itself: budgets, approvals, and rollback enforced the moment a trigger becomes an action — before money moves or data changes.
Per-agent identity, policy gates, atomic rollback, and an immutable audit trail give engineering and security teams proportional control, calibrated to each agent’s autonomy.

Catch it before the incident, not after.

FAQ

Common questions

Billing, governance, and how AgentDelivery fits alongside other webhook tools.

How is AgentDelivery different from Hookdeck or Svix?

Hookdeck and Svix make sure your webhook is delivered. AgentDelivery governs what happens next — what an agent is allowed to do when that event arrives: budget and rate gates, human approval on sensitive actions, atomic rollback, per-agent identity, and a full audit trail, all enforced before the action runs. Delivery is the on-ramp; the action is the risk.

What counts as an event for billing?

Every inbound webhook admitted through AgentDelivery counts as one event. Policy-blocked requests do not count. Overage above the plan quota is billed at €12 per block of 25,000 events. See pricing for plan quotas.

Which plan includes the agent guardrails?

All of them. Every paid plan — Starter €19, Core €49, Scale €89 — includes the full Agentic Event Delivery Guardrails: agent SDKs with per-agent credentials, budget/rate/content gates, human-in-the-loop approvals, atomic bundles with rollback, and OTEL traces. Plans differ on volume (events, endpoints, retention), not on safety.

Can I replay failed webhook deliveries?

Yes. Every event retains its full payload and delivery history. Replays are instant, idempotent, and recorded in the audit trail — the deterministic execution layer prevents duplicate side effects even when the same event is replayed multiple times.

Which webhook sources does AgentDelivery support?

Any HTTP source. First-class integrations exist for Stripe, GitHub, and Shopify (including signature verification). Custom sources connect via a standard ingest URL with HMAC or Bearer authentication. See the quickstart for setup.

Full FAQ →

Live System Status

Public operational summary with no tenant identifiers.

API Loading

Ingest Loading

View full status