The Autonomy-Learning Maturity Model (A x L)

As of August 2026.

Why another maturity model

Ladders for agent autonomy already exist, and this guide builds on them rather than pretending otherwise. Google DeepMind published autonomy levels 0-5 framed as a human-interaction paradigm and a deliberate deployment choice (Morris et al., arXiv Nov 2023, ICML 2024). Feng, McDonald, and Zhang published L1-L5 defined by the user's role, from operator to observer, with a proposal for autonomy certificates (arXiv Jun 2025). Hugging Face researchers published levels of agency defined by who controls program flow, with an argument against the fully autonomous top level (Mitchell et al., arXiv Feb 2025). Salesforce ships an L0-L4 agentic maturity model keyed to orchestration scope (Apr 2025 [vendor]). All descend from Sheridan and Verplank's ten levels of automation (1978) and the SAE driving levels.

What none of these models carries is a learning axis: across the field we surveyed, learning capability appears in almost no maturity model at all. Yet in enterprise practice, whether an agent can improve, and under what governance, determines the operating model around it more than its autonomy does: a fixed-policy agent is configured; a learning agent is managed. This model therefore has two dimensions.

Axis 1: Autonomy (A0 to A5)

Who executes the work, and what role the human plays. The A-levels are an enterprise operating-model ladder, applied per workload; they are closest in spirit to Feng's user-role levels and DeepMind's interaction paradigms.

Level	Name	Work execution	Human role	Required controls
A0	Manual	Humans with conventional tools	All work	n/a
A1	Assisted	Copilots suggest; humans execute	All decisions and actions	Usage policy
A2	Delegated tasks	Agents execute bounded tasks on request	Review every output	Task scoping, output review, audit
A3	Supervised autonomy	Agents run routine workflows end to end	Approve irreversible actions, handle exceptions	Approval gates, evals, budgets, kill switch
A4	Managed autonomy	Agent teams run routine operations to SLOs	Manage exception queues, supervise	Eval-gated change, concern register enforced, supervisor dashboards, designed oversight capacity (see the gate below)
A5	Governed lights-out	Full routine automation in scoped domains	Set policy, own accountability, audit outcomes	Proven eval maturity, continuous assurance, regulator-ready evidence, oversight capacity measured in production, not designed on paper

Level: A0
Name: Manual
Work execution: Humans with conventional tools
Human role: All work
Required controls: n/a

Level: A1
Name: Assisted
Work execution: Copilots suggest; humans execute
Human role: All decisions and actions
Required controls: Usage policy

Level: A2
Name: Delegated tasks
Work execution: Agents execute bounded tasks on request
Human role: Review every output
Required controls: Task scoping, output review, audit

Level: A3
Name: Supervised autonomy
Work execution: Agents run routine workflows end to end
Human role: Approve irreversible actions, handle exceptions
Required controls: Approval gates, evals, budgets, kill switch

Level: A4
Name: Managed autonomy
Work execution: Agent teams run routine operations to SLOs
Human role: Manage exception queues, supervise
Required controls: Eval-gated change, concern register enforced, supervisor dashboards, designed oversight capacity (see the gate below)

Level: A5
Name: Governed lights-out
Work execution: Full routine automation in scoped domains
Human role: Set policy, own accountability, audit outcomes
Required controls: Proven eval maturity, continuous assurance, regulator-ready evidence, oversight capacity measured in production, not designed on paper

The bottling plant of the vision chapter operates at A3 to A4. Two cautions on A5, both from the evidence: a published argument exists against building fully autonomous agents at all (Mitchell et al., 2025), and production deployments in 2026 cluster at low autonomy levels (Gartner, 2025-2026). A5 is a governed end-state that few workloads should reach, entered per domain and never declared enterprise-wide.

Axis 2: Learning (L0 to L3)

Where and how the agent improves, and under what governance. This axis is the model's contribution, introduced by this guide.

Level	Name	How the agent improves	Governance
L0	Fixed policy	It does not; behavior changes only by redeployment	Standard change management
L1	Curated learning	Humans update its ground truth, heuristics, and prompts offline	Versioned content, review before release
L2	Governed learning	The system proposes improvements (rules, retrieval, routing) that promote through evals and human gates	The learning flywheel: trace, evaluate, curate, promote, roll back; promotion criteria explicit
L3	Continuous learning	Online memory and adaptation inside guardrails, with promotion of durable learnings through the L2 gate	Memory governance (provenance, TTL, erasure), drift monitoring, freeze switch

Level: L0
Name: Fixed policy
How the agent improves: It does not; behavior changes only by redeployment
Governance: Standard change management

Level: L1
Name: Curated learning
How the agent improves: Humans update its ground truth, heuristics, and prompts offline
Governance: Versioned content, review before release

Level: L2
Name: Governed learning
How the agent improves: The system proposes improvements (rules, retrieval, routing) that promote through evals and human gates
Governance: The learning flywheel: trace, evaluate, curate, promote, roll back; promotion criteria explicit

Level: L3
Name: Continuous learning
How the agent improves: Online memory and adaptation inside guardrails, with promotion of durable learnings through the L2 gate
Governance: Memory governance (provenance, TTL, erasure), drift monitoring, freeze switch

Three rules connect the axes:

Autonomy is promoted by earned trust, and learning maturity is the evidence. An agent moves up A-levels when demonstrated reliability justifies it (consistent with Anthropic's graduated-trust guidance, Aug 2025 [vendor]), and demonstrated reliability at scale requires at least L2: without governed learning, reliability can only be observed, not improved. As a practical rule, this guide holds that A4 and above should require L2 or better.
The teammate region. The vision chapter's claim, "without a learning loop you have a fixed-policy agent, not a teammate," lands here: L0-L1 agents are automation you configure; L2-L3 agents are teammates you manage, with a named accountable human governing what they learn.
Autonomy is capped by oversight capacity, not by model capability. A4 and above additionally require a designed and instrumented oversight capacity. The rule and its evidence are below. Added August 2026 after the R13 operating-model research; it is the same shape as rule 1, and for the same reason: a level you cannot supervise is a level you have not reached.

The oversight-capacity gate on A4 and above

This gate exists because of the single strongest negative finding in the guide.

No credible published human-to-agent supervision ratio exists. A major vendor introduced "human-agent ratio" as a management metric in 2025 and published no numbers; its 2026 successor still publishes none. Every figure in circulation for how many agents one person can supervise is an extrapolation, not a citation. An A-level ladder that quietly assumes such a ratio would be resting on nothing.

What the adjacent measurement does support:

Supervisory capacity follows the fan-out relation: capacity is neglect time divided by the sum of interaction time and wait time. Wait time is the term architects forget and the one that usually binds, because it is set by queueing and escalation design rather than by the agent.
The only directly measured capacity in a comparable supervisory task fell between four and six units, in a low-stakes simulation. It is a floor for intuition, not a target.
Published ratios across domains span a thirty- to fiftyfold range, driven entirely by task, autonomy level, and consequence. There is no domain-independent number to find.
Consequence-weighted domains regulate below one to one. Nuclear control rooms require multiple licensed operators per unit with continuous presence at the controls.
The process industries supply the only standards-embedded workload figure: more than one alarm per operator per minute is treated as very likely unacceptable, and more than a hundred alarms in the first ten minutes of an upset is expected to lead operators to abandon the system.

The gate, stated operationally. A workload may be claimed at A4 or above only when all four hold:

The oversight capacity is calculated from the fan-out relation with wait time included, using measured interaction and wait times from the workload itself.
The capacity target is expressed as a burst rate, never as a daily or hourly average. The standards committee that governs alarm load holds that a per-ten-minute rate cannot validly be converted into a longer-window figure, and removed its per-day metric precisely because averaging destroys the meaning. Any "N agents per reviewer" target set as a daily average will be met on the dashboard and violated in every burst.
Supervision load, intervention rate, escalation mix by trigger, and wait time per item are instrumented in production (R12 carries the mechanics).
Verification is cheap by construction. If checking the agent's work costs a meaningful fraction of doing the work, the capacity calculation is fiction regardless of what the telemetry says.

Calibrated oversight is the expected shape, not a regression. The evidence on trust progression is counterintuitive and worth stating inside the model: as operators gain experience with an agent they grant broader standing permission and intervene more often, not less. Rising intervention rates alongside expanding autonomy are a sign the model is working. A programme that reports expanding autonomy with falling intervention should be investigated before it is celebrated.

What the gate is not. It is not a third axis. Oversight capacity is a precondition on the autonomy axis in the same way governed learning is, and modelling it separately would imply enterprises can trade autonomy against oversight, which is exactly the trade this gate exists to forbid.

Using the model

Per workload, never enterprise-wide. An enterprise holds a portfolio of (A, L) positions. The autonomy level is a deliberate deployment choice, not the maximum the technology allows (a principle borrowed from DeepMind's levels work).
Movement is earned, and reversible. Promotion between A-levels follows demonstrated reliability against explicit eval criteria; demotion follows incidents. Feng et al.'s autonomy-certificate idea is a useful institutional form for this.
The controls column is the contract. Each A-level names the minimum control set; the guide's cross-cutting concern register (identity, observability, traceability, grounding, impersonation, sovereignty, privacy, safety, cost, resilience) applies at every level and tightens as A rises.
Anti-pattern: agentwashing the ladder. Claiming A3 or higher for systems that are scripted automation with a chat front end. The test is behavioral: does routine work flow to the agent by default, and does a human approve only exceptions and irreversible actions?

Sources

Morris et al., Levels of AGI, arXiv Nov 2023 / ICML 2024. https://arxiv.org/abs/2311.02462
Feng, McDonald, Zhang, Levels of Autonomy for AI Agents, arXiv Jun 2025. https://arxiv.org/abs/2506.12469
Mitchell, Ghosh, Luccioni, Pistilli, Fully Autonomous AI Agents Should Not Be Developed, arXiv Feb 2025. https://arxiv.org/abs/2502.02649
Salesforce, Agentic Maturity Model, Apr 10 2025. https://www.salesforce.com/news/stories/agentic-maturity-model/ [vendor]
Anthropic, framework for safe and trustworthy agents, Aug 4 2025 [vendor]
Kasirzadeh and Gabriel, Characterizing AI Agents, arXiv Apr 2025. https://arxiv.org/abs/2504.21848
Sheridan and Verplank (1978); Parasuraman, Sheridan, Wickens (2000); SAE J3016 (lineage)
Gartner, agentic AI maturity and adoption commentary, 2025-2026 (via press releases and licensed reprints)
Olsen and Wood, fan-out and neglect-time formulation of supervisory capacity (2004 onward); human-supervisory-control literature
EEMUA 191 and IEC 62682 alarm-management practice, including the removal of the per-day metric and the ten-minute burst rate
Full evidence base and the span-of-control finding: research/R13-operating-model/findings.md and sources.md

Source: synthesis/maturity-model.md in the evidence repository behind this site.