Home / ADLC / 5 · Five Autonomy Levels
ADLC · Slide 5 Autonomy Spectrum

5 Autonomous
Levels

From Suggest to Autonomous. A calibrated spectrum for deciding how much execution to delegate to an agent and how to progress safely across it.

ADLC 5 Autonomy Levels Suggest, Assist, Supervised, Bounded Autonomy, Autonomous Slide 5 · ADLC Framework · Author: Prashant Dhingra
← L1 · Human executes L3 · Co-piloted L5 · Agent autonomous →
The five levels

Autonomy is earned, not assumed

Each level defines the boundary between agent execution and human oversight. Start at the lowest viable level and progress only on demonstrated reliability.

Level
01
Suggest
Agent proposes · Human executes

The agent generates options, recommendations, and drafts but the human makes every execution decision and performs the actual work. The agent functions as an accelerated thought partner: it structures the problem, surfaces possibilities, and prepares material. Nothing is committed without explicit human action.

Code suggestions in IDE Draft PR descriptions Architecture options list Refactor recommendations
Agent autonomy
Agent Proposes solutions, drafts content, generates candidate approaches. Output is always advisory never committed.
Human Reviews every suggestion. Selects, rejects, or modifies. Performs all execution. The agent is an advisor, never an actor at this level.
Level
02
Assist
Agent executes each step · Human approves each step

The agent executes individual steps autonomously, but each step requires explicit human approval before the next begins. This is fine-grained delegation: the human retains full visibility and a veto at every micro-decision point. Suitable for tasks where the path is predictable but the stakes of any individual misstep are meaningful.

Step-by-step code generation Database migration scripts Incremental test scaffolding Multi-step debugging
Agent autonomy
Agent Executes each discrete step fully and presents the result for review before proceeding. No step is skipped or batched without explicit permission.
Human Approves each step before the agent continues. Can redirect, roll back, or stop at any point. The gate is per-step, not per-task.
Level
03
Supervised
Agent completes full task · Human reviews before merge

The agent works through an entire task end-to-end without step-by-step approval, then presents the complete output for human review before it is merged or committed. This is the co-pilot model: the agent handles all implementation and self-corrects during execution, but the human holds the merge gate and cannot be bypassed. The most widely adopted ADLC level in practice.

Full feature implementation Automated test generation Full module refactor Documentation complete pass
Agent autonomy
Agent Completes the full task including self-correction loops. Produces a final output ready for human review no intermediate check-ins required.
Human Reviews the complete deliverable before it is merged or deployed. The merge gate is human-held and cannot be automated away at this level.
Level
04
Bounded Autonomy
Agent runs freely inside guardrails · Human reviews at gate

The agent operates with broad freedom within a set of human-defined guardrails established at Stage C (Planning). It can execute, iterate, test, and adjust without per-step or per-task approval until it reaches a stage gate. At the gate, a human reviews the cumulative output before work proceeds. Guardrails define the operational envelope; the gate is the accountability checkpoint.

Autonomous sprint execution Continuous test & fix loops Multi-file refactoring Performance optimisation
Agent autonomy
Agent Runs freely within guardrails: executes, tests, iterates, and self-corrects across the full scope of the stage without waiting for approval at each action.
Human Reviews at the stage gate only not during execution. Sets and owns the guardrails that define the agent's operating boundary. Can intervene if a guardrail breach is detected.
Level
05
Autonomous
Agent executes and self-deploys · Human edits afterwards

The agent executes the full task and deploys to production within hard technical limits (not human-set guardrails). Human involvement is post-facto: review, edit, and correction happen after deployment, not before. This level is only appropriate for tasks with very low error cost, high reversibility, and a proven track record of agent reliability at that specific operation. Full pre-deployment human review is replaced by fast rollback capability.

Auto-deployed hotfixes Canary release management Dependency updates Config drift correction
Agent autonomy
Agent Executes and self-deploys within hard technical limits. Monitors its own output post-deployment and triggers rollback if criteria are breached.
Human Reviews and edits after deployment. Sets and enforces the hard limits that constrain agent action. Owns the rollback decision if the agent's self-assessment fails. Full accountability remains human.
At a glance

Comparing all five levels

When to use each level and what the human touch-point looks like at each.

Level Agent scope Human touch-point Use when
L1 · Suggest
 Proposals only no execution Every decision and action New task type; low agent familiarity
L2 · Assist
 Executes each step Approves each step before next Predictable path; meaningful per-step stakes
L3 · Supervised
 Completes full task end-to-end Reviews before merge Proven task type; RICE-A all pass
L4 · Bounded
 Free within guardrails, full stage Reviews at stage gate only High agent reliability; guardrails well-defined
L5 · Autonomous
 Executes and self-deploys Edits and monitors post-deploy Low error cost; high reversibility; proven track record
Moving up the spectrum

How to progress safely

Level advancement is governed by evidence, not elapsed time or team confidence. These three principles define safe progression.

📊

Measure before you move

Track agent accuracy at the current level across at least 10 comparable tasks. A ≥95% success rate with no high-severity escapes is the threshold for considering progression to the next level.

🧪

Re-run RICE-A at each level

A task that passes RICE-A at Level 3 may not pass at Level 4 if the scope of autonomous action widens the blast radius. Re-evaluate every criterion when considering a level increase.

🔁

Levels are per task-type, not per agent

An agent trusted at L4 for unit test generation may only be at L2 for authentication logic. Autonomy is earned per task category, not granted globally to an agent or team.

⚡

Regressions trigger level reduction

A high-severity failure at any level immediately returns that task type to Level 3 (Supervised) as the ceiling. Reinstatement to higher levels requires a fresh evidence period.

🚦

L5 requires institutional sign-off

Moving to full autonomous operation is not a team-level decision. Level 5 requires architectural review, rollback infrastructure validation, and explicit risk acceptance by a named human authority.

📋

Document the level assignment

Every task type's current level assignment should be recorded explicitly in the team's ADLC operating document. Undocumented autonomy is ungoverned autonomy and ungoverned autonomy is risk.

"The goal of the autonomy spectrum is not to reach Level 5 everywhere it is to reach the right level for each task. A team operating thoughtfully at Level 3 is safer and more productive than one rushing to Level 5 on untested ground."

Prashant Dhingra, Agentic Development Lifecycle Framework