The delivery engine for every channel, and the agents that answer back.
Providers rate-limit. Networks drop. Fire-and-forget means gone forever.
Absorb the burst. Pace their limits. Send twice, deliver once.
Back off, retry, recover. Out of attempts, park for replay. A bad address never retries at all.
A million promos never delay a login code.
…and the world answers back.
The engine listens back — opened means the reminder never fires.
at-least-once · idempotent · observable end to end
Delivered was just the beginning.
04 · agents
It answers. You stay in control.
05 · the edit
Prompt edits are deploys.
Watch one travel.
A prompt is code. One line of it changes what the agent does.
Scenarios replay on the real pipeline. Tools are asserted, the reply is scored. The model grades, the code decides.
Every push runs the suite against the real agent. A push that breaks it cannot merge.
Your customers edit prompts too. Theirs are graded before Save commits — the check warns, it never blocks.
Every save is a version. A new one runs on a slice of real conversations, judged beside the old — an unjudged control arm is not a control.
The easy turns run on a cheaper model — trusted to talk, never to act. Escalation is law, not judgment.
Two gates around the brain: what the agent will discuss, and what a reply may never say. A prompt can be argued with. A gate cannot.
One flooding customer is throttled politely and free. Everyone else never notices, and the record stays truthful.
One edit. Seven stations. Nothing reached a customer that hadn't earned it.
06 · proof
Don't take my word. See the proof.
evals · replayed, not mocked
run: refund-scenarios · 3 of 3 passed
observability · every turn traced
turn #14 · 3.2s
handoff · it knows its limits
cost · flat, not climbing
tokens per turn · last 8 turns
cache_control · the prefix is cached, not re-paid
TOTAL production-ready AI agents in your channels
observability evals guardrails knowledge memory cost human handoff