Agentic engineering

Software engineering that delegates bounded implementation work to AI agents and evaluates the result with repository rules, tests, security controls, runtime evidence, and change history.

Agentic engineering is software development in which agents can inspect a codebase, edit several files, run commands, and respond to failures inside a defined environment. A request for an API endpoint can become a patch, focused tests, and a report of the observed behavior. The surrounding engineering system still defines what counts as correct: public contracts, architecture boundaries, permissions, test suites, deployment gates, and rollback paths. An agent may own a bounded change from investigation through verification without owning the product's goals or receiving unrestricted access.

Builder example

The useful shift is from supervising each generated line to designing work that can be checked. A low-risk, well-specified patch may pass through automated tests and merge policy with little intervention. A change involving payments, private data, or unclear product intent may need specialist or owner review. The review route follows the evidence and potential impact, while the repository keeps the same maintainability and security standards regardless of who wrote the code.

You ask an agent to refactor a billing form. It changes validation, display copy, and database writes in one pass.

Constrain the task, inspect the change surface, run contract tests, and route ambiguous billing behavior to the appropriate specialist.

Common confusion: Agentic engineering does not mean that every output needs manual line-by-line approval, and it does not mean an agent receives unlimited authority. It means delegation, permissions, evaluation, and recovery are explicit enough for the risk of the change.