Methodology · Iteration Managementemergingverified

Five-Level Agent Progression

also known as agent autonomy ladder, L0-L5 agent levels, Bornet five levels

Applies to: agentmulti-agent-systemcoding-agentbrowser-agent

Tags: maturity-ladderautonomy-levelsplanning

A six-step ladder for how capable an AI agent is. The steps run from Level 0, fully manual, up to Level 5, fully autonomous, with rule-based, automated, and workflow stages in between. A team picks a target step based on how risky the task is and how much value it brings. To climb a step, the lower step's safety checks must be in place and the agent must pass its tests. This rates the agent as a whole, not one action at a time. Teams use it to agree on what 'done' means before they start building.

Methodology process overview

stateDiagram-v2 [*] --> L0: Manual baseline L0 --> L1: rules + automation in place L1 --> L2: ML triggers added (IPA) L2 --> L3: agentic workflow with planning L3 --> L4: semi-autonomous with HITL gates L4 --> L5: fully autonomous, evals + guardrails proven over months L5 --> L4: regression demotes one step pending fix L4 --> L3: regression demotes one step pending fix L3 --> L2: regression demotes one step pending fix note right of L3 Each promotion requires: - prior-level guardrails in place - eval set passing at threshold - documented evidence end note note right of L4 Skipping steps (e.g. L1->L4) ships demos that break on production drift end note

Intent. Place an agent on a six-step capability ladder. This makes the target level of independence, and the safety checks needed to reach it, clear before anyone builds.

When to apply. Use this early, when you are scoping an agent project. It helps most when people disagree on how independent the system should be, or when you are planning a roadmap that spans several quarters. It also works as a look-back tool: 'we said L4, we shipped L2.5'. Don't apply it once building is underway and the per-action gate (crawl-walk-run) is the right tool. The ladder is for planning, not for runtime control.

Example scenario

An insurance carrier is scoping an agent for claims processing across three task families: first-notice-of-loss intake (high volume, low risk per call), document-completeness checks (medium volume, medium risk — wrong call delays payouts), and settlement-amount calculation (low volume, high risk — wrong call costs real money or under-pays customers). The CTO wants L5 'fully autonomous claims agent'. The chief actuary wants L1 'rule-based with human review on everything'. The framework is used to align them. They map each task family to a candidate level: FNOL intake at L3 (agentic workflow, since multimodal documents and unstructured speech require reasoning), document-completeness at L2 (IPA — ML triggers on existing OCR pipeline), settlement calculation at L1 (rule-based — actuarial tables don't need an agent, and regulators won't accept opaque reasoning). The CTO and actuary both sign off; the target is explicitly mixed across tasks rather than one number for the whole agent. For FNOL intake — the most ambitious — they list the L2 preconditions: OCR pipeline at 98% field-extraction accuracy on the existing form set, exception-routing rules in place, observability into every call. The L3 advancement criteria are: planner-eval passing at the threshold the agent-platform team published, and 60 days of L2 operating with the published exception rate. Construction proceeds one level at a time. Six months in, the team reviews and confirms FNOL has reached L3 with the named evidence; the actuary asks whether settlement could move to L2 given new explainability tools, and they re-open that question. The ladder is a planning instrument, revisited as portfolio risk shifts, not a static org chart.

Inputs

Task portfolio — The set of tasks the agent will handle. Each task is marked with its risk and its value.
Stakeholder constraints — Limits on how independent the agent can be. These come from regulators, the organisation, and how much risk people will accept.
Prior-level evidence — Test results and past incidents from the level the system runs at today.

Outputs

Target level declaration — The target step on the L0 to L5 ladder, set per task or per system, with the reason for it.
Advancement criteria — The named safety checks, tests, and live signals required to climb one step.
Capability roadmap — A step-by-step plan to move from the current step to the target step.

Steps (6)

Map tasks to candidate levels
For each task, find its natural step. The steps are manual (L0), rule-based (L1), ML triggers on existing automation (L2), an agent that plans (L3), semi-autonomous with human gates (L4), and fully autonomous (L5). Let the task's risk and value drive the choice, not technical excitement.
Declare a target level per task or system
Write down the target. Get explicit sign-off from stakeholders. Mixed targets are normal. A support agent may be L4 for FAQ and L2 for refunds.
Define prior-level guardrails as preconditions
For each step, list the safety checks and tests that must already pass at the step below before you consider climbing. Moving from L4 to L5 usually requires months of logged human acceptance at the published bar.
usesHuman-in-the-Loop Approval Queue
Build to one level above current
Climb one step at a time. Skipping steps, such as jumping from L1 to L4, tends to ship demos that break once real traffic shifts.
Gate advancement on evidence
Climb only when the named safety checks are in place and the test set passes the bar. Write down the evidence. If the numbers drop at a higher step, the system moves down one step until it is fixed.
usesShadow Canary
Revisit target level on portfolio shifts
When a task's risk or value changes, such as a new regulation or a new product line, set the target step again. The ladder is for planning, not a permanent label.

Framework-specific instructions

Pick a framework and generate a framework-targeted rewrite of this methodology's steps.

Choose framework

AI-generated for Agent Development Kit (ADK) (Google) — verify against official docs.

Principles

The autonomy level is a planning choice. It is driven by task risk and value, not by technical capability alone.
Climb one step at a time. Skipping steps ships demos that production breaks.
Higher steps need the lower steps' safety checks already in place.
The ladder is for aligning stakeholders, not for runtime control. Pair it with per-action gating at runtime.

Five-Level Agent Progression

Methodology process overview

Steps (6)

Map tasks to candidate levels

Declare a target level per task or system

Define prior-level guardrails as preconditions

Build to one level above current

Gate advancement on evidence

Revisit target level on portfolio shifts

Framework-specific instructions

Principles

Known failure modes (3)

Related patterns (4)

Related compositions (2)

Related methodologies (2)

Sources (2)

Provenance