A prompt in, a product out.
Skills, tools, connectors, policies, evals and memory, generated together and running in Playground within minutes. Refine any part by hand or by prompt.
Describe the work, the systems and the rules. Get a governed agent that does the job, asks before it acts, and ships as one file you own. Deploy it to your customers under your brand.
Open-source runtime. Any model. One WASM file, yours forever.
Describe the agent you want to sell
A Xero bookkeeping agent for small property managers. Read bills and receipts from the shared mailbox, code them to the right property and landlord, raise rent invoices, chase arrears, draft replies to tenants and landlords. Never post or send without approval.
Nothing here is hand-written. Real runtime, replaying a recorded session. Nothing is sent, nothing is posted.
Read the invoice: Acme Plumbing Ltd, 12 Sep 2026, £312.00 inc VAT, “boiler service, Flat 2”. Matched the supplier and the property. Not a duplicate of the £312.00 bill from March: different invoice number and date.
How I worked this out · 1 document read · 3 tools · 1 Xero lookup · 2.1s
Every run is a deterministic log of inputs and effects. Replay it in Studio, on your server, or in this page. Same result every time.
Skills, tools, connectors, policies, evals and memory, generated together and running in Playground within minutes. Refine any part by hand or by prompt.
Every generated agent carries hooks that check arithmetic, provenance and read-back around every model call and every write, workflows that make big jobs resumable, and approval gates as code.
One WASM file holds the agent and its rules. Host it with us under your brand, or run it in your customers’ own tenants. Any model: Anthropic, OpenAI, Azure, Ollama.
Why these agents do not make things up
A prompt is not a guarantee. So every agent the generator produces comes with deterministic hooks around every model call and every write, and with workflows that turn big jobs into resumable plans with human gates. You can read them, edit them, and watch them fire.
Code that runs at fixed points: after the model answers, before anything is written, after Xero confirms. It can pass, fix, send the model back with a reason, block, or ask you.
fn after_model(ctx, out) {
let b = out.bill;
if b.net + b.vat != b.total {
return retry("net + VAT does not equal total; re-read the document");
}
if b.vat_rate == 20 && (b.net * 0.20 - b.vat).abs() > 0.01 {
return retry("VAT is not 20% of net");
}
if !ctx.document.text.contains(b.invoice_number) {
return retry("invoice number not present on the document");
}
pass()
} Larger jobs as declared plans: steps, fan-out, approval gates, a checkpoint after every step so a run survives a restart, and compensation when a late step fails after an early write.
Checkpoint after every step. Compensation: a bill that fails in Xero after approval notifies the approver; nothing is voided silently.
Sell it on
Each customer gets their own organisation, connectors, approvals and audit trail. You set the price. One dashboard across all of them. We bill you per organisation.
For your data-protection reviewer
No credentials inside. The only hosts it may call are listed in the manifest. Signed at export, verified on load. Tenant data stays in your tenant and your chosen model region.
"connectors": [
{ "name": "xero", "hosts": ["api.xero.com", "identity.xero.com"] },
{ "name": "m365-mail", "hosts": ["graph.microsoft.com"] } ],
"secrets_embedded": 0,
"policies": ["xero-write/approval", "email/approval", "bill/duplicate-block", "payments/none"]Studio with the generator, Playground and evals. Deploy to your customers under your brand and set your own price. We charge you for the factory and per running organisation.
Runtime, host drivers, embed element and pack spec are open source. Build a pack with Markdown, YAML and Rhai. Embed the same agent in your product.
$ npm i @atomharness/host @atomharness/embed $ dotnet add package AtomHarness.Host $ atom init my-pack && atom run
Package names are the plan, not yet published. The waitlist is where the first invitations to the repository go.
Waitlist
Studio is in private preview. Tell us who you are and what you would build, and we will invite you as capacity opens up.
No list. No newsletter. One email when there is something to try, and a way to be forgotten.