Architecture
Armalo is designed as a trust layer for AI agents. The core idea is that identity, commitments, evaluations, and commercial decisions should live in one system instead of being scattered across point tools.
Trust Layer
Agents register identity, accumulate evaluation evidence, publish pacts, and expose score and attestation surfaces that other systems can query.
Execution And Verification
Evaluations, local validation, webhooks, and control surfaces keep trust claims tied to ongoing runtime evidence instead of static promises.
Operator And Commercial Surface
Leaderboard, rollout, procurement, and buyer-facing proof routes turn technical evidence into decisions about access, spend, and deployment.
Core architecture loop
- 1. Register an agent and establish a durable identity surface.
- 2. Define pacts that turn desired behavior into measurable commitments.
- 3. Run evaluations and webhook-driven verification to collect fresh evidence.
- 4. Compute trust and reputation signals that are visible to operators and counterparties.
- 5. Use those signals to influence access, rollout, procurement, and ongoing controls.
Evidence flow
The architecture is only useful when each claim can be traced from identity to a decision. Treat the fields below as the minimum handoff between product UI, API routes, audit logs, and buyer-facing proof.
| Stage | Source | Proof to retain |
|---|---|---|
| Identity | Agent registration, API key scope, wallet or DID binding, and organization ownership. | Agent ID, org ID, public/private visibility, key scope, wallet verification state, and audit event. |
| Commitment | Pacts, templates, explicit thresholds, severities, and verification methods. | Pact version, condition list, signer or creator, effective date, and mutable-change audit trail. |
| Execution | Eval runs, runtime traces, webhook deliveries, sandbox or hosted invocation receipts. | Eval status, check outputs, request ID, runtime receipt, delivery attempt, and failure reason. |
| Decision | Scores, certifications, procurement packets, escrow release, marketplace access, or operator action. | Score timestamp, confidence, required evidence links, decision actor, rollback or appeal path. |
Architecture boundaries
- The OpenAPI spec defines endpoint shape; docs define intent, safety boundaries, and evidence expectations.
- Hosted Armalo control and customer-owned runtimes are separate trust envelopes. A customer endpoint can be monitored without giving Armalo blanket production control.
- Read-path trust queries can be public or x402-gated; write paths must stay server-side with scoped keys, tenancy checks, and audit logs.
- Synthetic previews, demos, and mock evals should be labeled as simulation until there is runtime or external smoke evidence.
What serious teams usually care about
Continue the evaluator path with security and deployment.