From Suggest to Autonomous. ADLC defines exactly how much work can be delegated to an AI agent, and how humans retain accountability as autonomy increases.
An agent may be Level 4 for test generation but only Level 2 for authentication logic. Trust is earned per task type.
Higher autonomy changes review timing, not ownership. Humans still define boundaries, risk posture, and release authority.
Teams advance levels only after repeated success, RICE-A validation, and clear rollback or intervention paths.
The goal is not universal Level 5. The goal is the right level for each task, based on reversibility, inspectability, constraints, error cost, and proven reliability.
The agent provides options, suggestions, drafts, and recommendations. The human makes all execution decisions and performs the work. This is the safest starting point for new task types.
The agent performs one step at a time, but the human must approve before the next step begins. This gives the team speed while preserving tight control.
The agent completes a full task from start to finish, including self-correction, then submits the final deliverable for human review before merge or deployment. This is the common ADLC operating mode.
The agent works freely within boundaries set during planning. The human does not approve each step, but still reviews at the stage gate and intervenes if boundaries are breached.
The agent completes the task and deploys within hard technical limits. Human review happens after deployment. This only fits low-error-cost, highly reversible, proven task types with strong rollback capability.
Each level changes where the human touchpoint happens: before every action, before merge, at the gate, or after deployment.
| Level | Agent scope | Human touchpoint | Use when |
|---|---|---|---|
| L1 · Suggest | Proposals only | Every decision and action | New task type |
| L2 · Assist | Executes each step | Approves each step | Predictable path, meaningful stakes |
| L3 · Supervised | Completes full task | Reviews before merge | RICE-A passes |
| L4 · Bounded | Runs inside guardrails | Reviews at stage gate | High reliability, clear boundaries |
| L5 · Autonomous | Executes and self-deploys | Reviews post-deploy | Low error cost, proven track record |
Teams should advance autonomy only when the agent has proven reliable in the same task category, under the same constraints, with measurable success.
Track performance on repeated similar tasks before increasing autonomy.
Reassess reversibility, inspectability, constraints, error cost, and autonomy fit as scope grows.
A high level for tests does not imply a high level for security, auth, architecture, or release decisions.
High-severity failures should reduce the task type to a lower autonomy level.
Full autonomous execution requires architectural review, rollback validation, and explicit risk acceptance.
Every task type should have a documented current autonomy level and gate expectations.
The aim of the autonomy spectrum is not universal Level 5. The aim is the right level for each task.
Prashant Dhingra, Agentic Development Lifecycle FrameworkMove from autonomy levels into stage gates and operational governance for agentic development.