How it works

Code is the primary tool

The model decides what to compute. Code does the computing.

Most platforms answer a question about data by putting the data in front of the model. Cantora writes a program, runs it, and hands back the answer.

The worked example

“How did this week compare with last week?”

the program Cantora writes1,023 records read

// Runs once, in a fresh sandbox, under a read credential
// scoped to this Tenant by the provider itself.
const orders = await commerce.orders.list({
  placedAfter: "2026-07-14",
  placedBefore: "2026-07-28",
});

return Object.entries(Object.groupBy(orders, byIsoWeek))
  .map(([week, weekOrders]) => ({
    week,
    orders: weekOrders.length,
    revenue: sum(weekOrders.map((o) => o.total)),
  }));

what the model reads612 bytes

[
  { "week": "2026-W29", "orders": 512, "revenue": 184230.55 },
  { "week": "2026-W30", "orders": 511, "revenue": 201884.10 }
]

A size-bounded one-shot TypeScript program, validated, run in a fresh sandbox under finite time and resource limits.

The model never sees the 1,023 records. It sees the answer, and explains it.

The sandbox is bounded by credential authority and lifetime, not a network filter, and a Calculation is an inspectable envelope, not a network audit.

What the person sees

Revenue analysis AgentTest

How did this week compare with last week?

Revenue is up +$17,653.55

Read from

Shopify
Stripe
Orders

2026-W30

Orders511
Revenue$201,884.10
Change+$17,653.55

The answer lands where the question was asked: the figure, the trend it sits on, and the week's numbers behind it.

Underneath it sits the Calculation — the exact program, the input, the authority it ran under, and the 612-byte result the model actually read.

The same question, two ways

Context stuffing

Load both weeks into the context window and ask the model to do the arithmetic. You pay for every token, wait for them, and trust a model to sum a column.

A code operation

Query, aggregate, return. Cost tracks the answer rather than the data, and the arithmetic is deterministic.

What that saves

Loading the records into the context window100%

Every one of the 1,023 order records is read by the model, and billed, to answer a question about two totals.

Running it as one code operation6.3%

The model reads a 612-byte result. The records are read inside the sandbox, by the program, and never enter the context window.

93.7%fewer total tokens on this workload, counted across the whole Turn rather than the tool result alone. The median across all six measured workloads is 95.2%, ranging from 75.3% to 99.3%.

95.2%Median reduction in total tokens across six measured structured-data workloads, counted across the whole Turn rather than the tool result alone.

75.3–99.3%The spread across those six workloads. The saving is stated as a median with its range, never as a constant.

612 bytesWhat 1,023 order records become in the commerce workload, which measures 93.7% on this basis.

The saving grows with the data — exactly where context-stuffing gets most expensive. Cantora absorbs usage inside the fee, within your written envelope, so every token removed is Cantora's saving to earn and your price to keep flat.

Cost is only the first consequence

Cost

The bill tracks the size of the answer, not the size of the dataset.

Correctness

Aggregation happens in code, not in a model's arithmetic. A sum is a sum.

Auditability

The program, input, authority, and result are retained. A reviewer inspects real logic, not a model's account of it.

The difference

An explanation is not evidence. A retained Calculation is.

Behavior is versioned, evaluated, and reversible

Any change to a model, prompt, Tool, policy, or budget creates a new candidate that has to earn its release, evaluated whole against the Live baseline rather than by a model score.

Promotion is a controlled rollout and rollback is a release, not a redeploy. A provider's new model is qualified against your Evaluations before it carries traffic.

See what the first two weeks produce, what it costs to run, or where the security boundaries sit.

Get started

Scope your first agent

Bring one bounded workflow and its systems. Check readiness, or describe it now.

Scope your first agent