Sandboxed agents. Receipts they cannot forge.
This page is for your technical due diligence. Four mechanisms carry the trust model, and each one is described as built, not as aspiration.
Each task runs in its own machine.
A task does not share a machine with other tasks, other companies, or other customers. The machine is created for the task and torn down after it.
Network egress is default-deny. Today the machine can reach exactly one destination: the Peragam control service. Even DNS is blocked. Provider calls happen through that gateway, never from the machine directly.
- One machine per task
- Default-deny egress
- Torn down after the run
The agent cannot grade its own work.
Verification runs as a separate machine check against the task type’s Definition of Done: a deploy must answer with a live URL, a plan must parse against a versioned output contract.
Receipts are written by the platform after that check. The agent that did the work has no write path to them, so a receipt cannot be forged from inside a task.
A receipt shows only evidence the verifier actually observed. Missing evidence renders as absent, never as invented proof.
- Independent verifier
- Platform-owned write path
- No invented evidence
Brand-risk actions wait for you.
Every company has an approval policy: a default tier, plus per-task-type overrides you control.
Outreach, social posts, and ad spend sit behind those gates. A gated task holds as blocked until you clear it, and clearing it is recorded.
- Per-type policy tiers
- Blocked until cleared
- Approvals recorded
Settlement is atomic. The stop is hard.
The credit charge and the verification result settle in the same step. A passed check charges the task’s credits; a failed check charges zero. There is no state where a failed task has billed you.
An admission guard stops new work before your balance can go negative. Domain purchases carry their own pre-authorized spend cap in dollars.
Customer payments run on your own Stripe account. Funds settle to you directly; the platform takes a flat 5% application fee on charges agents create there.
- Zero credits on failure
- Hard credit stop
- Your Stripe account, flat 5% fee
Due-diligence questions
Can an agent mark its own task as done?
No. Done is a state a task can only reach through a passing verification receipt, and receipts are written by the platform after an independent check. The agent has no write access to receipts or task status.
What can an agent reach from inside its machine?
One destination: the Peragam control service. Egress is default-deny and DNS is blocked, so everything else is unreachable. The machine never holds provider keys.
What happens if agents try to overspend?
An admission guard refuses new work before the credit balance can go negative, and domain purchases are capped by a pre-authorized dollar limit you set.
Who holds my customers’ payments?
You do. Payments run through your own Stripe account and settle to you directly. The platform takes a flat 5% application fee on charges agents create there.
Can I see why a task failed?
Yes. A failed receipt shows what the verifier observed, such as a non-200 HTTP status, along with the zero-credit charge.
Bring your hardest questions.
Join early access and check the receipts yourself.