Agentic artifacts as code¶

Definition¶

Treating agentic artifacts as code is the practice of placing every input that shapes an agent's behaviour — system prompts and prompt templates, tool manifests, policy configuration, memory schemas, evaluation datasets and model settings — under the version control, semantic diffing, review and formal change-approval discipline normally reserved for application source, on the grounds that these artifacts are production dependencies whose uncontrolled edits are a leading cause of agent failure.

Explanation¶

In a conventional service the deployable surface is code plus configuration. In an agentic service that surface expands to a set of text and JSON/YAML artifacts that are trivially editable, frequently live in a vendor console outside the repository, and have no compiler or type checker to catch a regression. A tool manifest behaves like a package dependency: adding or altering a tool changes what the model can decide to do, and because a tool's output is appended to the next prompt, the change propagates into subsequent reasoning steps rather than staying local. Prompt edits interact unpredictably with model upgrades and data changes; the source names prompt drift as the dominant production failure mode and attributes the majority of production agent failures to uncontrolled prompt modification. The resulting discipline is concrete — store the artifacts in Git, require formal approval to change them, ship changes through progressive delivery such as A/B testing a prompt revision, and wire automatic rollback to behavioural metrics rather than to error rates, because a drifted agent generally does not throw. Runtime prompt-control platforms are compatible with the discipline only insofar as they still record and pin versions. The source is an InfoQ practitioner article and its failure-rate attributions cite vendor blog posts rather than controlled studies, so read the practice as accumulated operational experience.

Key Properties¶

Behaviour-determining surface: prompts, tool manifests, policies, memory schemas, eval sets, model configuration
Tool manifests act as dependencies — a change alters capability and propagates into later reasoning through returned output
Prompt drift is claimed as the dominant production failure mode for agentic systems
Rollback triggers must watch behavioural metrics, since drift raises no error
Runtime prompt control is compatible with the discipline only if versions remain pinned and recorded

Relationships¶

Golden trajectory regression testing — supplies the behavioural baseline that makes rollback of these versioned artifacts decidable, since a drifted prompt raises no error to trigger on
Agent harness — names the same surface from the architecture side — the harness is precisely what these versioned artifacts configure

Applications¶

Bringing prompts and tool manifests out of a vendor console into the repository with review gates before an agent goes to production; adding semantic diffs and approval to prompt changes so a behavioural regression is attributable to a specific commit.

Sources¶

https://www.infoq.com/articles/prompts-to-production-playbook-for-agentic-development/