ARYN.

Overview

ARYN (Autonomous Agent Infrastructure) is an infrastructure project for managing the complete AI agent lifecycle, from configuration and evaluation to execution and result review. Its focus is to make agent actions controllable, behavioral changes testable, and outcomes traceable.

The system connects a management interface, an API, operational storage, and runtime adapters. Developers can manage agent versions, define access boundaries, compare evaluations, and review actions that require approval.

Problem and objectives

An agent can produce a convincing answer while making decisions that are difficult to inspect. Changes to prompts, models, or tool access can also alter its behavior in ways that a short demonstration will not reveal.

ARYN addresses this through three objectives:

• Define an agent's authority before execution.
• Evaluate configuration changes against explicit scenarios and criteria.
• Connect results, approvals, and execution history to the agent version involved.

This approach supports agent development and operation where human review, access control, and a consistent decision history are required.

System architecture

The web interface communicates with the ARYN API. Core checks requests before forwarding approved execution to the Hermes adapter. Model access passes through 9Router so that model selection and availability can be checked at the service boundary.

The main responsibilities are:

• Core: user identity, organization and project scope, access policies, action approvals, usage, and auditing.
• Agent Factory: blueprints, configuration versions that become immutable after publication, project assignments, and rollback.
• Bench: evaluation scenarios, baseline comparisons, and regression checks.
• Brief: supporting evidence, information provenance, and the handling of contradictory findings.
• Relay: incident investigation and recovery on registered, approved targets.

These boundaries keep the interface and execution runtime subject to the policies enforced by the governing services.

Agent workflow

1. A developer prepares a blueprint and saves a configuration version.
2. The candidate agent runs through Bench.
3. Evaluation results are reviewed before approval.
4. An approved version is published and assigned to a project.
5. Core checks permissions, model availability, and usage limits before execution.
6. Results and audit information are retained for review.

Approval is associated with the configuration and action reviewed. Configuration changes require the corresponding review. Request keys help prevent duplicate execution when the same request is submitted again.

Engineering decisions

Preserve configuration versions. Agent changes create a new version, keeping earlier evaluations and outcomes connected to their original configuration.

Evaluate before promotion. Bench distinguishes passing, failing, and unverified results. A critical regression against the accepted baseline can prevent publication.

Enforce authority in the service layer. Model output and a tool's capability description do not automatically authorize an action. Sensitive operations remain subject to policy checks and approval.

Make failures observable. Unknown model availability stops the relevant operation. Model identity, output, usage, and audit context support investigation.

Development outcomes and evaluation

Development covers agent configuration management, Core persistence and governance, Bench evaluation workflows, and Hermes runtime integration. Evaluation combines deterministic checks with comparisons against an accepted baseline.

The testing approach spans service contracts, PostgreSQL integration, access behavior, the interface, and browser workflows. The main concerns are consistent authority, version traceability, and failure handling throughout the lifecycle.

Status and next steps

ARYN remains in development. Operational readiness requires further validation of identity providers, hosted database operations, TLS, reverse proxies, runtime networking, and model integration in the target environment.

The next priorities are to complete that validation, expand evaluation scenarios, and strengthen recovery procedures while retaining human review for sensitive actions.

Code and documentation

• ARYN project code
• Studio workflow documentation