AI agents can't be
trusted with money.
We're the layer that fixes it.
Exactly-once, human-approved, compliance-gated execution for autonomous agents — running live on Cloudflare. We prove it on outbound; the same engine runs any high-stakes, long-running AI workflow.
The business case
Elite senders win by configuring better, not sending more — but automation at scale has real downside. This engine makes the expensive failures impossible by construction.
Configuration beats volume — and durability protects the revenue.
Cold email lives or dies on relevance: personalizing deeply from real research, keeping to ≤80 words and one soft CTA, sending from an authenticated domain, and honoring opt-out and suppression. But when an AI writes and sends unattended, a single mistake — a wrong claim, a double-send, a bounce — costs real money. So the engineering standards aren't academic: each one defends a revenue line.
A crashed worker re-emails the prospect → torched domain deliverability.
A retried step emails the same prospect twice → burned lead, spam risk.
High-value account contacted without an AE → burned relationship.
Send without opt-out / suppression → CAN-SPAM / GDPR exposure.
Live command center
This talks to the real deployed engine. The scorecard is computed by the live /score endpoint; the console creates a genuine durable workflow instance you approve to completion.
Use cases
One durable engine, one authoring model (Python or YAML → one IR), many outbound motions.
Prospect → one personal email
Pull prospects from Genesis, research each site, draft one grounded email, and send it exactly-once after your approval. The default deployed workflow.
A human before every send
Each lead pauses at human.approve and hibernates for free until you sign off — no polling, no cost while waiting, nothing sent unattended.
Hibernate for free
Instances park at step.sleep / wait-for-event for hours or weeks at zero compute cost — the primitive behind any delayed or approval-gated motion.
Start on any signal
Kick off a durable run from your own systems via POST /runs when a signal fires — a funding round, a job change, a form fill — and reach out inside the window.
Regulated regions
GDPR / CASL motions where the terminal compliance guard hard-blocks any send missing opt-out, suppression or frequency capping.
API-driven orchestration
Create durable runs from your own systems via POST /runs, deliver approvals via /events, and poll status — all documented below.
The workflow, in detail
The deployed 6-step outreach workflow. Each node is a durable, individually-retriable step whose result is journaled — so a crash replays completed work instead of re-running it.
Operation economics
What one fully-durable, AI-personalized, human-approved run actually costs — and how little of the platform budget it touches. Estimated at published Cloudflare Standard rates; confirmed against the real run.
The trust layer for AI that touches revenue
Every company is racing to put AI agents on revenue-critical work. Almost none can trust them to act. That trust gap is the opportunity — and the moat.
Autonomy you can trust with your pipeline, your domain, and your reputation.
AI can already write and send the outreach. What it can't do — and why "AI SDR" demos stall before production — is be trusted to act unsupervised on things that cost real money: a double-send torches deliverability, an unapproved enterprise blast burns an account, a non-compliant send is legal exposure. We built the durable-execution layer that makes autonomous revenue operations safe, and proved it on the hardest surface: outbound. The same engine runs sales and fundraising.
The bottleneck was never sending
It's the research, personalization and compliance needed to earn a reply at scale. Human SDRs are ~$70–100k, slow to ramp, and inconsistent. AI can do the work — but only if you can trust it near your revenue.
A category forming right now
Sales-engagement & GTM tooling is a multi-billion-dollar market, being re-platformed by AI. Three curves just converged: capable/cheap LLMs, durable-execution primitives, and edge economics — so millions of long-lived, human-approved AI workflows finally run for cents.
A platform, not a script
Exactly-once execution, fail-closed human approval, compliance guards, full cost/observability — the trust layer rivals skip. One IR with three authoring surfaces (SDK, YAML, visual builder), plus integration leverage: partners plan, we execute.
Software-grade margins, at any scale
One fully durable, AI-personalized, human-approved run costs ~$0.0002 (≈ $0.23 per 1,000, $226 per 1,000,000) — see the live cost dashboard above. Usage-based revenue on near-zero marginal cost means the margin structure scales with customer volume, not headcount. Every figure on this page is computed by the live engine — nothing is mocked.
A working platform, not a deck
A production durable engine deployed on Cloudflare — plus a Python SDK (@workflow), a YAML loader, a drag-and-drop visual builder, and the GENESIS Capital integration, all compiling to one IR. Real workflows run end-to-end (retries, hibernation, human approval, exactly-once) with a live cost dashboard and multiple independent security & quality reviews applied. (Public demos simulate sends; authenticated campaigns on a verified domain send for real via Resend.)
Outbound is the beachhead
The same engine is the execution layer for any high-stakes, long-running, human-approved AI workflow — collections, onboarding, renewals, procurement, compliance ops. The durable trust layer for AI agents that touch money and reputation. Outbound is where we prove it and win the wedge.
Engineering standards & best practices
The contract the platform holds itself to — each derived from adversarially-verified research and enforced in code, not aspiration.
Architecture
Three layers, one durable engine. Authoring is plural; execution is singular.
Interface
This page — a live console over the API — plus a visual builder that round-trips the same declarative spec.
Authoring SDK
Flow decorators and declarative YAML both compile to one validated intermediate representation — code, YAML and the builder are the same object at runtime.
Durable engine · Cloudflare Workflows (Python)
Each instance is a SQLite-backed Durable Object running the IR as memoized steps: retries, hibernation, wait-for-event, and a terminal compliance guard.
API reference
The same endpoints this page uses. Write endpoints are rate-limited and validate input.
| Route | Method | Purpose |
|---|---|---|
| / | GET | this console (static asset) |
| /api · /health | GET | service info + workflow fingerprint |
| /spec | GET | the compiled workflow IR |
| /score | POST | run the guru scorecard on a config JSON |
| /runs | POST | create a durable workflow instance · rate-limited, spec-validated |
| /runs/{id} | GET | instance status & output |
| /runs/{id}/events | POST | deliver an event (AE approval) · rate-limited |