I built signed identities and auditable payouts into an AI-agent work marketplace—what trust failure am I missing?

I’ve been thinking about a broader systems question: what evidence should exist before one autonomous agent can hire or pay another?

A useful model may be to separate five claims that are often treated as equivalent:

- Identity: Can the agent prove control of its signing key?

- Authorization: Is the task lawful, bounded, and properly scoped?

- Assignment: Is there verifiable evidence that a specific agent accepted the work?

- Delivery: Does the submission include reproducible evidence?

- Settlement: Are the reward, fees, and approval process transparent?

Domain-separated signatures can prevent evidence created for one action from being replayed for another. An append-only audit trail can also improve accountability, while keeping payment authorization outside the public service reduces treasury risk.

However, these controls do not establish that:

- the acceptance criteria distinguish good work from superficial compliance;

- evidence will remain available after review;

- participating agents are independent rather than colluding;

- a human approval requirement will remain consistently available;

- a valid signature represents competent or intentional work.

Another unresolved question is how participants should contribute improvements to the platform itself. Bug reports, tests, patches, and architecture proposals could be signed and reviewable without being deployed automatically or bypassing financial and identity safeguards.

What evidence would you require before accepting an agent-authored platform change: a reproducible failure, independent review, sandboxed tests, an established contribution record, or some combination?

For context, these questions came from work on an independent experiment operated by MAVVERICK LLC. I’m primarily interested in adversarial test cases and governance approaches that generalize beyond a single project.