The autonomy contract

Two axes, applied per workload

In short: Each piece of work gets two settings, how independent the agent is and whether it may learn, both chosen on purpose and reversible.

The first axis is the autonomy levelAutonomy level (A0 to A5). How much an agent may do without a person in the loop, from manual (A0) to governed lights-out operation (A5). Set per task, not per agent.: who executes the work, and what the human keeps hold of. Its levels are A0 manual, A1 assisted, A2 delegated tasks, A3 supervised autonomy, A4 managed autonomy, and A5 governed lights-out. The second axis is the learning levelLearning level (L0 to L3). How an agent's behaviour is allowed to change over time, from fixed (L0) to continuous learning inside guardrails (L3).: whether the agent improves, and under what control. Its levels are L0 fixed policy, L1 curated learning, L2 governed learning (the learning flywheelLearning flywheel. The controlled loop that makes agents better over time: collect evidence, test, approve, roll out in stages, and pull back what fails.), and L3 continuous learning inside guardrailsGuardrail. A filter that uses an AI model to catch bad inputs or outputs. Useful as an early warning, but not reliable enough to be the thing that grants or denies an action.. The two axes move independently. Every workload sits somewhere on the grid deliberately, and movement is earned and reversible.

foundation · matrix

The A×L maturity matrix

Autonomy and learning are independent; higher is not automatically better.

Dimension	L0 Fixed policy	L1 Curated learning	L2 Governed learning	L3 Continuous learning
A0 Manual	·	·	·	·
A1 Assisted	·	·	·	·
A2 Delegated tasks	·	Common starting point	·	·
A3 Supervised autonomy	·	·	·	·
A4 Managed autonomy	·	·	Oversight gate	·
A5 Governed lights-out	·	·	·	Rare and domain-scoped

A0 Manual

A1 Assisted

A2 Delegated tasks

• L1 Curated learning:Common starting point

A3 Supervised autonomy

A4 Managed autonomy

• L2 Governed learning:Oversight gate

A5 Governed lights-out

• L3 Continuous learning:Rare and domain-scoped

Diagram description: Six autonomy levels A0 through A5 crossed with four learning levels L0 through L3, with governed scaling emphasized in the middle of the matrix. The matrix crosses A0 Manual, A1 Assisted, A2 Delegated tasks, A3 Supervised autonomy, A4 Managed autonomy, A5 Governed lights-out with L0 Fixed policy, L1 Curated learning, L2 Governed learning, L3 Continuous learning. Annotated cells are A2 Delegated tasks by L1 Curated learning: Common starting point; A4 Managed autonomy by L2 Governed learning: Oversight gate; A5 Governed lights-out by L3 Continuous learning: Rare and domain-scoped.

Reviewed 2026-08-20·Sources:A×L maturity model

The controls column is the contract

In short: Each step up in independence comes with a fixed list of controls, and that list, not the level name, is the agreement.

Three rules connect the axes. A4 and above require L2 or better. At that level reliability must be improvable, not merely observable. A4 and above require proven oversight capacity. A level you cannot supervise is a level you have not reached. Without a learning loop you have a fixed-policy agent, not a teammate. Fixed policy is the right choice for many workloads, as long as it is named honestly.

Readiness gates autonomy

In short: Six areas of readiness are each scored from 0 to 3, and the weakest area, not the total, decides how much autonomy is safe.

Six dimensions are each scored from 0 to 3: absent, ad hoc, managed, operated. Data readiness includes ACL-aware retrievabilityACL-aware retrieval. Retrieval that checks the requester's permissions before returning information, so an agent cannot surface documents its user could not open., meaning retrieval that respects each person's access rights. Integration readiness covers governed APIAPI. A defined way for software to talk to other software. Agents reach systems through governed APIs rather than through screens. coverage (application programming interfaces, the defined ways software talks to software), toolTool. A function an agent can call to do something real: look up a record, send an email, run a query. Tools are where agents touch your systems. exposure, and event availability. Identity readiness covers agents as first-class principals, on-behalf-ofToken exchange (on-behalf-of). Swapping one credential for another so an agent acts with the permissions of the person who asked, not with its own broader powers. flows in which the agent acts with the requesting person's permissions, and secrets never held by agents. Operational discipline covers observabilityObservability. Being able to see what agents did, step by step, from records the agent cannot alter., offline and online evalsEval (evaluation suite). A set of test tasks with known right answers, written by people who know the work, used to check an agent before and after any change., and incident practice. Governance and value discipline covers intake with kill criteria (agreed conditions for stopping), per-run cost visibility, and the two-estate viewTwo estates (metered and licensed). Metered: agent traffic that flows through your own gateway and can be fully governed. Licensed: agents inside vendor products that run on the vendor's controls, governed only through their settings and exports.. The two estates are agents behind your own gatewayGateway. The single door every agent request passes through on its way to a tool or model. Because everything goes through it, it is where rules are enforced and activity is recorded. and agents inside vendor products. The sixth dimension is operating capacity. The profile governs, never the total:

foundation · profile

Readiness is a profile, not a total score

The weakest relevant dimension constrains the workload’s autonomy ceiling.

Data readiness

0–3

Quality, ownership, permissions, provenance

Integration readiness

0–3

Governed APIs, tools, and event access

Identity readiness

0–3

ID2 principals, delegation, secrets, sponsor

Operational discipline

0–3

Observability, evals, incidents, drift

Governance and value discipline

0–3

Intake, risk, value, and per-run cost

Workforce and operating model

0–3

Owners, supervision capacity, and change

Diagram description: Six-part readiness profile covering data, integration, identity, operations, governance and value, and workforce without collapsing the dimensions into one score. The scorecard calls for Data readiness (Quality, ownership, permissions, provenance), Integration readiness (Governed APIs, tools, and event access), Identity readiness (ID2 principals, delegation, secrets, sponsor), Operational discipline (Observability, evals, incidents, drift), Governance and value discipline (Intake, risk, value, and per-run cost), Workforce and operating model (Owners, supervision capacity, and change).

Reviewed 2026-08-20·Sources:Readiness assessments

Identity gaps cap autonomy without blocking a start. A workload can begin at A1 while the identity chain is built.

The oversight gate, as a burst rate

In short: No credible supervision ratio is published, so oversight capacity comes from queueing maths, budgeted per ten minutes, never daily.

No credible published human-to-agent supervision ratio exists from any source. The vendor that coined the metric has published no number in two annual editions. What transfers instead is queueing arithmetic and alarm-management standards:

foundation · profile

Design oversight for bursts, not averages

Exception demand is uneven, so safe autonomy depends on recoverable surge capacity.

Open evidence gap

1. 1

   Routine

   Exceptions arrive within staffed capacity.

2. 2

   Weak signal

   Several agents encounter the same upstream fault.

3. 3

   Burst

   Correlated exceptions exceed the normal queue.

   Gate · Capacity threshold

4. 4

   Degrade safely

   Pause, narrow permissions, or return work to people.

5. 5

   Recover

   Clear backlog and verify before restoring autonomy.

Measure arrival shape, handling time, and recoverability for each workload.

Diagram description: Timeline of low routine exception demand interrupted by a correlated incident burst that exceeds ordinary human review capacity and triggers degraded mode. The sequence contains 5 stages: 1, Routine: Exceptions arrive within staffed capacity.; 2, Weak signal: Several agents encounter the same upstream fault.; 3, Burst: Correlated exceptions exceed the normal queue., followed by the Capacity threshold gate; 4, Degrade safely: Pause, narrow permissions, or return work to people.; 5, Recover: Clear backlog and verify before restoring autonomy.. Important boundary: Measure arrival shape, handling time, and recoverability for each workload.

Reviewed 2026-08-20·Sources:A×L maturity model

• Capacity follows the fan-out relation. Fan-outFan-out. How many agents or tasks one supervisor can watch at once, which shrinks as waiting times grow. is how many agents one supervisor can watch at once. It equals neglect time (how long an agent can safely run unattended) divided by interaction time (how long a person needs per intervention) plus wait time. Leaving wait time out overstates capacity by up to 67 percent, and by 36 percent even in exception-only designs.
• Budgets are burst rates, never daily averages. Oversight budgets are set as burst ratesBurst rate. Demand measured at its busiest moments rather than averaged over a day. Oversight capacity is planned against bursts. per ten minutes (demand at its busiest, not averaged), never as daily averages. The alarm-management standards hold that a per-ten-minute rate cannot validly be converted to a longer window. They removed their per-day metric because averaging destroys its meaning. A daily-average target will be met on the dashboard and violated in every burst.
• Expect calibrated oversight, not a ratchet. Calibrated oversightCalibrated oversight. As operators gain experience they grant broader standing permission and also step in more often, rather than supervising less. means that with experience, standing permission broadens and intervention rates rise. In measured deployments, auto-approval roughly doubled while interrupt rates nearly doubled alongside it. Rising interventions with expanding autonomy is the system working. Falling interventions with expanding autonomy is the thing to investigate.

Measure these, as R13, the operating model layer, sets out. Track the intervention rate and its trend, the mix of escalations by trigger, wait time per item, and verification cost per review. Add periodic blind checks against self-reported figures.

The research behind this page

• The maturity model
• Readiness assessments
• Operating model findings