The governed engineering environment is the substrate of machine-actionable constraints within which AI coding agents do their work: the harnessing, types, lints, schemas, deployment processes and analyses that the authors of Cheap Code, Costly Judgment: A Case Study on Governable Agentic Software Engineering describe as necessary for agent productivity. In that paper it is the place governance is relocated to — the engineering governance thesis holds that agentic velocity is sustainable only when governance lives in the environment rather than resting on human review, because otherwise human attention becomes the throughput bottleneck. The paper uses agentic governance in the operational sense it attributes to recent agent-governance work: the mechanisms by which autonomous agents can be kept productive while the cost of their failures is held within acceptable bounds, where a mechanism may be either a control that detects or contains failures or an architecture that eliminates them by construction.
Usage
The environment is not designed up front but accumulates through Governance Conversion, and the paper reports it growing large relative to the product it governs. In the case studied, the support apparatus measured 1.16 MLOC against 420 KLOC of production code — 2.75× — made up of static analyses (238 KLOC), dynamic analyses (405 KLOC), agent-referenced documentation (247 KLOC), agent infrastructure (110 KLOC) and tooling (162 KLOC). Concretely it included 577 project-specific static analyses alongside commodity analyzers (Roslyn/.NET, Pyright/Python, ESLint/TypeScript), more than 13,000 C# test methods, roughly 1,500 Python test files, over 300 property-based tests and 90 fuzz harnesses.
The authors catalogue 41 representative mechanisms in ten families, divided by what they constrain. Agent governance families cover governance-doc controls (making agent-facing project rules enforceable rather than conventional), context and dispatch (brief-linting, dynamic context injection, role-typed dispatch), agent observability (agent registries, typed event buses, sentinel commits), resource mediators (serializing shared host resources across concurrent worktrees) and incorporation gates (pre-commit hooks, merge-train batching, staged deploy gates). Product governance families cover canonical seams, validation and conformance, static and dynamic analyses, provenance and attribution, and repair vocabulary — bounding remediation actions to closed, typed sets.
Mechanisms tend to be layered rather than singular: when a failure recurred across one mechanism, the response was to add a complement, and the deepest such stack in the case reached 14 complementary mechanisms around orchestrator/sub-agent interactions. The paper also argues that soft, probabilistic encodings — templates an orchestrator should follow, conventions stated in documentation — saturate under agentic velocity and must be mated to deterministic controls such as compiler-checked typed enumerations and analyses that gate commits, because at sufficient volume a low-probability harness violation becomes a certainty.
Related Terms
Governance Conversion, Process Models for Agentic Software Engineering, Harness Engineering, Guardrails, Agent Hooks