Skip to main content

What is Agentic Development Governance?

Agentic Development Governance (ADG) is a systematic approach to governing autonomous or semi-autonomous AI agents that participate in software development activities.

Canonical definition​

The formal definition of ADG and related terms lives in Canonical Definitions.

This page explains the problem framing: why ADG exists and what it is trying to prevent.

What ADG focuses on​

ADG focuses on the engineering and operational reality of building software with agents:

  • preserving coherence under autonomous change (code, schema, infra, docs)
  • enforcing boundaries (auth, data access, tenancy, environments)
  • ensuring auditability and replayability of agent actions
  • preventing drift that can “work today” while becoming brittle or unsafe over time

What ADG is (and is not)​

ADG is​

  • a governance layer for agentic coding and agentic DevOps
  • a discipline for maintaining architectural intent under high-velocity change
  • an approach to preventing “works now / breaks later” failures

ADG is not​

  • a generic “AI ethics” program
  • model governance or data governance
  • a code assistant
  • a linter
  • a single CI check
  • a compliance checklist divorced from engineering reality

Why ADG is needed now​

For decades, software safety relied on a stable assumption:

  • humans wrote the code
  • humans reviewed the code
  • humans understood the code

Agentic development breaks that assumption:

  • agents can generate massive surface area quickly
  • the resulting system can “work” while hiding severe flaws:
    • tenancy leakage
    • authorization ambiguity
    • hidden coupling between environments
    • undocumented side effects
    • schema drift and migration damage
    • contradictions between docs and runtime behavior

ADG is the missing layer that makes agentic software auditable, enforceable, and maintainable.

Core outcomes ADG targets​

  • Transparency: actions are visible and explainable
  • Accountability: actions are attributable and reviewable
  • Controllability: humans can pause/stop/override agent actions
  • Compliance: policies are enforced at critical boundaries and gates
  • Continuous improvement: the system learns from incidents and near-misses