Agents as code: bringing software discipline to agent operations

Agents are becoming production software

Enterprise agents increasingly contain prompts, instructions, tools, workflows, structured outputs, permissions, and business rules. Treating that configuration as an informal collection of copies inside a user interface creates familiar software problems: unclear source of truth, inconsistent versions, difficult rollback, and fragile releases.

Agents need the operational discipline enterprises already apply to code.

A repository-backed source of truth

BasilAOS supports agent definitions and templates managed through connected repositories. Administrators can maintain multiple sources, synchronize them on demand or on a schedule, choose the template used for publication, and publish one agent or a fleet of agents.

This changes the operating model. A template is no longer a hidden configuration copied by hand. It becomes a versioned artifact that can move through review, release, and change control.

Separate template from instance

Enterprise rollout often requires a common definition with different permissions, data scopes, tools, or local configuration for each team. BasilAOS distinguishes template-level and instance-level controls so organizations can standardize the agent while preserving the access and operating boundaries of each deployment.

Publishing is only one part of operations

Agents also need observability and lifecycle management. Teams should know which version is active, which repository supplied it, what tools and models it called, how users interacted with it, where failures occurred, and whether an unused agent should be retired.

Repository-backed templates, version management, call records, run logs, and idle-agent governance create the foundation for that discipline.

Toward agent CI/CD

The long-term pattern resembles software delivery:

• define and review the agent as a versioned artifact;

• test instructions, tools, structured outputs, and permission boundaries;

• publish to selected instances;

• observe production behavior and exceptions;

• feed approved corrections into the next version;

• roll back or retire safely when required.

The model may remain probabilistic, but the system around it does not need to be informal. Agents as code makes enterprise AI easier to govern, reproduce, and scale.