Learn / Start here

One task, end to end

Follow one agent task from mandate to challenge, then switch off a part and see which guarantee disappears.

Integration test · not yet performed · Reviewed 3 October 2026

One task, end to endIntegration test · not yet performed

Switch off a part. Watch a guarantee disappear.

The task. A researcher asks an AI agent to analyse a sensitive dataset within a budget, and to report back.

Parts of the architecture Identity proofsIdentity & VEIDAgent controlsDSEMACompute marketVirtEngineAttestationVirtEngineGovernance & challengeFoundation · network · DSEMA

Person → agent

The researcher grants a scoped mandate: which data, which actions, what budget, until when. A purpose-specific proof shows they hold the authority to grant it.

Holds. A mandate tied to an accountable, eligible person, revealing only the attribute the purpose needs.

Published design

VEID consent and credential records are in public source; DSEMA's authority-granting process is published design. The two are not connected.

x/veid consent records

What could an independent reviewer answer from the final record?

6 of 6 answerable in the design

AnswerableWho authorised this, and were they entitled to?
AnswerableDid the agent stay inside its mandate?
AnswerableWho supplied the compute, on what terms?
AnswerableWhat environment ran it, and who could see the data?
AnswerableCan the outcome — or the rule — be contested?
AnswerableWas more personal information exposed than the task needed?

Every question is answerable only when every part is present. That is the claim the integration has to prove: no single program provides it.

No end-to-end run of this task has been performed. The components marked “in public source” have not been integrated with DSEMA, which is a published design. This walkthrough states what the integration would have to demonstrate. Sources: VirtEngine source, DSEMA specification, Foundation constitution.

Authorise (Person → agent). The researcher grants a scoped mandate: which data, which actions, what budget, until when. A purpose-specific proof shows they hold the authority to grant it. With every part: A mandate tied to an accountable, eligible person, revealing only the attribute the purpose needs. Without it: The mandate traces only to a key. Either nobody accountable stands behind it, or the person must disclose far more than the task requires. Status: VEID consent and credential records are in public source; DSEMA's authority-granting process is published design. The two are not connected.

Check (Agent controls). Before acting, the agent's request is checked against three independent gates: authority, resource ceiling, and containment. With every part: Only actions inside the mandate proceed. Being able to pay is not permission to act. Without it: Nothing compares the action with the mandate before execution. The ability to pay becomes the only gate. Status: Published DSEMA design with patent-pending claims for chain-verified instruction gating. Bosun's review gates are a narrower, operating analogue for code changes.

Acquire (Agent → providers). The agent places an order. Independent providers bid; a lease is formed and the budget is held in escrow. With every part: Resource terms anyone can inspect, from a provider the agent did not have to trust in advance. Without it: The agent depends on one operator's private terms. The resource ceiling has no shared record to be enforced against. Status: Market, escrow, and provider modules are in public source. The network is not live; public TestNet is planned for January 2027.

Execute (Provider hardware). The workload runs in a hardware-isolated environment. Attestation reports what environment ran and who could reach the data. With every part: Evidence of the execution environment, so sensitive data is not exposed to the provider's operators. Without it: The researcher must take the provider's word for what ran and who could see the data. Status: SGX, SEV-SNP, and Nitro attestation verification is in public source. Attestation shows the environment, not that the output is correct.

Settle (Protocol). Usage is recorded, the provider is paid from escrow, and the record links mandate, gate decision, lease, and attestation. With every part: A record connecting what was authorised to what was spent, without storing personal documents. Without it: Payment happens, but nothing ties the spending back to the mandate that allowed it. Status: Usage, settlement, and payout paths are in public source. Linking settlement to a DSEMA gate decision is not implemented.

Challenge (Person → governance). The researcher disputes the bill or the result. Separately, anyone may contest the rule that allowed the action. With every part: A defined route to contest an outcome and, more slowly, to change the rules — with delay and human ratification. Without it: There is no route to contest the outcome or the rule. Whoever controls the system decides alone. Status: Billing dispute workflows are in public source. DSEMA's two-house amendment process is published design. The Foundation constitution is in force but binds only the Foundation.

Switching parts off requires JavaScript.

01 / Explanation

Why one task, not four programs

Each program can be assessed on its own: an identity design, an agent architecture, a compute market, an engineering tool. The research question is about what they do together. The clearest way to state that is a single task carried across every handoff: authorise, check, acquire, execute, settle, and challenge.

02 / Explanation

A guarantee lives in a handoff

An accountable mandate needs identity and governance. An enforced limit needs agent controls and a market where the limit binds. An auditable record needs attestation and settlement linked to the mandate. Remove any part and the others keep working, but the record no longer accounts for the whole task. That is the sense in which the combination could create something the parts cannot.

03 / Explanation

What has and has not been done

VirtEngine's market, escrow, attestation, and dispute paths and VEID's consent records are in public source, on a network that is not live. DSEMA's gates and amendment process are published design. Nothing connects DSEMA to VirtEngine today, and no end-to-end run has been performed. The walkthrough is a statement of what the integration would have to demonstrate, and the first question on the research agenda.

Check your understanding

Every part works on its own, but agent controls are switched off. What is lost?

Answer: Any check that the agent stayed inside its mandate. Payment, leases, and attestation still function. What disappears is the comparison of each action with the mandate before it runs, so the ability to pay becomes the only gate.

Next: People, machines, and economic agency →All guides

Purpose & people

Governance & transparency

Connected research

Technology programs

Research questions

Interactive guides

Learn & build

Constitution explorer

Records & updates

Learn & contribute

Contact & membership

One task, end to end

Switch off a part. Watch a guarantee disappear.

The researcher grants a scoped mandate: which data, which actions, what budget, until when. A purpose-specific proof shows they hold the authority to grant it.

What could an independent reviewer answer from the final record?

Why one task, not four programs

A guarantee lives in a handoff

What has and has not been done

Every part works on its own, but agent controls are switched off. What is lost?

One task, end to end

Explore the mechanism

Switch off a part. Watch a guarantee disappear.

The researcher grants a scoped mandate: which data, which actions, what budget, until when. A purpose-specific proof shows they hold the authority to grant it.

What could an independent reviewer answer from the final record?

Why one task, not four programs

A guarantee lives in a handoff

What has and has not been done

Every part works on its own, but agent controls are switched off. What is lost?