Agency
See how it worksDelegate first outcome
Menu
ProductProofSystemsStartSign inDelegate first outcome
example tasksdelegate any outcome
Agency

A computer
that acts.

Delegate outcomes, not prompts. Agency plans the work, searches the live web, reads the primary record itself, asks for approval when it matters, and leaves a verifiable receipt trail.

Delegate first outcomeSee how it works
signed approvalssigned receiptsstanding monitors
example receiptssigned & verifiable
Example receipts are cryptographically signed and verifiable.
live web researchprimary-source readsmetered spendsigned approvalsexecution receipts
What it is

Agency is a work computer that delegates outcomes under signed authority.

Chat answers. Assistants reply. Agency acts — and proves what happened.

Most AI products stop at text. Agency turns intent into executed, receipted work: live research briefs, Stripe payout audits, standing monitors on a pre-signed schedule — every run metered, approved, and verifiable offline.

The principle

Calm agency — powerful enough to act, constrained enough to trust, visible enough to supervise.

01

Delegate

Give the outcome. Agency decomposes the work into steps, tools, and checkpoints.

02

Observe

Watch the task surface evolve without babysitting every click or subtask.

03

Approve

High-impact actions pause for permission before spend, send, delete, deploy, or commit.

04

Verify

Every completed action can produce a signed receipt: who, what, when, cost, and result.

Accountable automation

The missing primitive is proof.

Autonomous work needs more than a chat transcript. It needs durable identity, bounded authority, and receipts that survive outside the session.

Identity

Who acted?

Agent identity persists across sessions, providers, tools, and devices.

Authority

Was it allowed?

Policy gates define what the agent can see, spend, send, change, or delegate.

Receipt

What happened?

Execution receipts create a verifiable audit trail for every delegated task.

A single receipt — one completed task, signed.
019dc500…a17dverifying…

Audited 1,284 Stripe payouts against the ledger for May 2026. Flagged 3 mismatches over $50 (total $214.37). Paused before exporting the report — held for your approval, nothing sent or changed.

browserstripe.payouts.readledger.read
evidence 328adafc698d… · recorded 2026-04-25
$ motebit-verify sovereign-receipt-stripe-audit.json
verifying…
verified in your browser · @motebit/verifier@1.12.1 · no server

An authentic, tamper-evident record of what the agent did — signed by the key its identity commits to, verified offline. It proves the record is genuine and unaltered, not that the world matched the claim.

Don’t trust this page?inspect the receiptverify at receipt.computer
A linked chain — approval to execution to completion, plus a refusal.
task 019dc500…0a01verifying…
approved the owner consented

You approved send_email · risk 2

· approver-signed · pinned
↓ same task · consent matches the exact call (args_hash)
ran the agent executed it

send_email — the exact approved call

· sovereign
↓ same task
completed the task finished

Sent the Q3 contract to the vendor after you approved the send.

· sovereign
denied separately, at the line

Attempted to send $4,800 to a new payee. Denied by policy. No funds moved.

· sovereign · a different task, refused by policy
verified in your browser · @motebit/crypto · no server

Two signers, one accountable action: the owner approved this exact call, the sovereign agent ran it, the task completed — and the approval’s hash matches the call that ran, so consent is bound to the execution, not a vague yes. A separate over-policy act was refused. All signed, all verified offline.

Don’t trust this page?approvalcalltaskrefusal
System capabilities

A calm surface for serious work.

Sources

Reads the record

Live web search, then the primary document itself — SEC filings, HTML, PDF — with verbatim extracts bound into the receipt.

Outcomes

Delivers signed work

Research briefs, payout audits, monitor verdicts — each a signed, offline-verifiable artifact, never just a chat answer.

Standing work

Acts on schedule

Monitors run on a schedule you pre-signed, hold the standing thesis, and record what changed — with nobody watching.

Controls

Asks when it matters

Approvals are explicit for irreversible, costly, sensitive, or externally visible actions.

Start

Put Agency to work.

Three outcomes are live today: reconcile Stripe payouts against your ledger, commission a research brief cited from the live web, or arm a standing monitor that re-checks a thesis on your signed grant and briefs you only when it moves. All end the same way — a signed, offline-verifiable record of what was asked, what was allowed, what was done, and who consented.

Delegate an outcome →free · read-only · nothing moves without your signed approval
The rest of the vision — triage, drafts, outreach — is the same proof, extending to new outcomes as they ship.
Read the proof model →