AgentDelivery

What it is

AgentDelivery is an agentic event-delivery layer with guardrails built in. It sits between whatever triggers an agent action — a webhook, a tool call, a scheduled job — and the action itself, so every consequential step passes a budget check, an approval gate, and an idempotency check before it runs, and can be rolled back if it shouldn't have.

It works with the webhooks and event flows teams already run. Nothing about the agents changes: AgentDelivery is the policy and delivery layer in front of them, not a new framework to rebuild on.

What's built

• Budget & rate ceilings — kill runaway loops and cost spikes before any action runs, not after the invoice arrives
• Human-in-the-loop approvals — pause any sensitive or irreversible action until a person says yes
• Idempotent delivery — a retry never double-fires, so no duplicate refunds, charges, or double sends
• Automatic rollback & immutable audit — undo a bad action and see everything every agent did, with a deterministic decision trace per event
• Multi-tenant from line one — tenant isolation, per-tenant rate limiting, and signing-key separation as architectural commitments, not retrofits
• HMAC signing, idempotency keys, and loop detection on every event path
• Plan model — Free, Starter (€19), Core (€49), Scale (€89), Enterprise — with quota dimensions for events, endpoints, events-per-second, and retention window; Stripe billing wired end-to-end
• Developer tools, quickstart, and an operational blog at agentdelivery.io

Why it exists

The first wave of agent tooling optimised for letting agents do more. The cost of that is now visible: agents that loop, agents that retry an irreversible action, agents that spend real money on a wrong inference. The missing layer is the one that decides what an agent is allowed to do, enforces it deterministically, and keeps a record.

AgentDelivery is that layer. It assumes the action might be wrong and makes wrongness cheap to catch and cheap to undo — a ceiling before the spend, an approval before the irreversible step, an audit trail after the fact.

Architecture posture

Self-hosted on OVH. No platform lock-in, no managed-cloud dependency that locks the company into a renewal cycle. The stack is deliberately conservative — Go backend, Postgres for state, OTEL for observability, standard tooling end-to-end.

Multi-tenancy is a Class 1 invariant from the foundation. Tenant isolation, per-tenant rate limiting, and signing-key separation are non-negotiable architectural commitments, not features added later.

Idempotency is enforced on every external write. Stripe charges, downstream deliveries, third-party callbacks — all of them retry, and the system has to be safe under retry without operator intervention. That same property is what makes agent actions safe to gate and replay.

Who it's for

• Teams running AI agents that can spend money or take irreversible actions, and need a ceiling and an approval step between the agent and the consequence
• Agentic systems and AI pipelines that need durable, observable, policy-controlled delivery between components
• Engineering teams whose webhook and event layer is currently held together with cron jobs, ad-hoc retries, and manual intervention when deliveries fail
• Founders building event-driven or agentic platforms who want a well-engineered guardrail layer rather than building one twice

Status

Live and operational on agentdelivery.io. Billing is wired, plans are priced, the product runs in production. What's missing is distribution — AgentDelivery is pre-revenue by choice, focused on getting the surface right before going wide.

The current focus is bringing on early users and design partners — particularly teams already running agents with real spend or irreversible actions, who want to influence the next layer of features in exchange for early-adopter pricing and a direct line to the builder.