Tiarne Hawkins
← Insights and frameworks
Tiarne Hawkins · Assurance framework

The Agentic Operations Journey

Seven modes, M0 to M6. What changes across the ladder isn't how capable the model is — it's how much access, autonomy, action and accountability the enterprise has handed over, and what has to be true before it does.

Most AI governance still classifies by technology: which model, which vendor, which data. That tells you almost nothing about exposure. Two teams can run the same model and sit four rungs apart, because one drafts emails with it and the other lets it move money. The ladder classifies by delegation instead — and delegation is what the assurance floor has to be set against.

The modes · a ladder from M0 to M6

Where a use case sits on the ladder changes what readiness has to ask of it. The left half is a productivity story — the model helps people work. The right half is an operations story — the system does the work, and the enterprise carries the consequence. The curve is steep at the top because M5 and M6 add authority, not just capability: the same failure that produced a bad draft at M1 produces a committed, compounding action at M5. Select a rung to see what it demands.

Complexity & risk ↑ProductivityOperationsRisk grows non-linearly. Each rung widens the Four A's.M0AutomationM1PersonalproductivityM2TeamproductivityM3Workflow agentM4End to endagentic workflowM5Multi agentsystemM6Autonomous ops
A use case's rung is set by what has been handed over, not by the model underneath it — and a use case can climb one without anyone re-approving it. That quiet promotion is what the registry exists to catch.

Select a mode: click or — or press play in the bar above

M3

Workflow agent

An agent executes a defined workflow — real tool calls, real writes, inside one bounded process.

Assurance floor

  • Pre-deployment behaviour testing on the workflow itself
  • Tool-call correctness verified, not assumed
  • Brokered credentials — no standing access to systems of record
  • Per-run approval trail that reconstructs the decision

Four A's posture

Access
scoped to the workflow's systems
Autonomy
executes a defined workflow
Action
reversible writes
Accountability
workflow owner

Looks like this in FSI

KYC document intake, dispute case assembly, control-evidence gathering. The first rung with real writes.

Fails like this

At the edges of the workflow definition — inputs nobody scoped, handled confidently.

Where assurance attaches

AI Range test sets mapped to TEVV and the Four A's; Evidence Pack from the first run.

Each mode, described

A mode is defined by what the enterprise has handed over, not by the model underneath it. The same model can operate at M1 in one workflow and M4 in another. Read each rung for three things: what it is, what newly leaves human hands, and who is left in the loop — the glyph on the right is the topology, not decoration.

What changes at this rung
Who is in the loop
personagentsystem stepexception-only or scripted

Read the glyphs down the column rather than one at a time: people leave the loop before agents enter it, and by M5 the person is attached to the system by a dashed line — present, but no longer in the path of the work.

The ladder read through the Four A's

The rung is a summary; the Four A's are what it summarises. They escalate independently, which is why two systems on the same rung can need different controls. High Autonomy with low Action is a research agent — wrong answers, contained. Low Autonomy with high Action is a narrow agent wired to a payments rail — right answers until it isn't. Accountability is the dimension that gets skipped, because it is the only one nobody has to configure.

AccessAutonomyActionAccountabilitydelegated authority increasesM0Automationfixed endpointsnone, scripteddeterministicwritesprocess ownerM1Personalproductivityindividual scopesuggests to oneuserhuman executesthe individualM2Teamproductivityshared teamcorpussuggests at scalehuman executesteam leadM3Workflow agentworkflow systemsruns definedworkflowreversible writesworkflow ownerM4End to endagentic workflowcross-system,brokeredplans and replanscustomer-visibleexception deskM5Multi agentsystembroad, agent toagentdelegates toagentscompoundingassign an ownerM6Autonomous opsenterprise,self-requestedsets ownobjectivesirreversible atscaleboard-level
Fill density encodes escalation within a dimension, not across dimensions. Modes are cumulative: a system at M4 has already cleared everything below it.

Place your own use case

Pick the line that best describes the system you have in mind on each of the four dimensions. The ladder places it for you — and the rung it lands on is the assurance floor it has to clear.

Access
Autonomy
Action
Accountability
M0
Automation

Answer the four questions above to place a real use case on the ladder.

What the harness actually does

Assurance isn't a review that happens beside the system. The harness is the controlled execution and evaluation layer wrapping the agent — every call brokered, every action gated, every run instrumented into evidence. It is not a policy document, not a model card, and not a review board: those describe intended behaviour, while the harness constrains actual behaviour at runtime. The distinction matters most from M3 up, where a prompt asking the agent to stay in scope is not a control and cannot be evidenced.

Assurance harnesscontrolled execution & evaluation layerTask / requesthuman or upstream agentAgent runtimeplans, calls, replansbrokeredscoped returnsTools & datasystems of recordPolicy gateallow · hold · blockCommitted actioneffect in the businessevery run instrumented — prompt, call, return, decisionTEVVTestEvaluateVerifyValidateEvidence PackAlignmentReadiness Score
Remove the harness and three things go at once: the broker between the agent and the systems of record, the gate between a decision and its effect, and the trace that makes any of it provable afterwards.

Products against the ladder

Continuous TEVV means the same discipline runs before deployment and after it. The products differ by where in the lifecycle they attach — and the sequence matters: you cannot test what you have not inventoried, and you cannot monitor drift without a pre-deployment baseline to drift from.

ProductAttaches atJobProduces
Use Cases & AlignmentM0 – M6Registry of every AI use case, scored for compliance risk and placed on the ladder at the mode it actually operates at — not the one it was approved at.Inventory, risk classification, mode assignment
AI Rangepre-deploymentBehaviour testing against adversarial and cohort personas: task success, tool-call correctness, multi-step hallucination, policy enforcement under pressure.Test results, failure modes, go / no-go
Peregrinepost-deploymentContinuous monitoring of live agents with dynamic risk scoring — catching drift, scope creep and quiet promotion up the ladder that pre-deployment testing could not anticipate.Live risk posture, drift and escalation alerts
Evidence PackthroughoutCollects what each stage emits into a reproducible, regulator-legible record of what was tested, what was found, and what was decided.Signed, replayable assurance evidence

Terms as I use them

Assurance harness
The controlled execution and evaluation layer wrapping the agent. Distinct from the agent runtime — the runtime decides, the harness constrains and observes.
TEVV
Test, Evaluate, Verify, Validate. Evaluation here is agent-specific behaviour testing, not classic model-risk challenger testing.
The Four A's
Access, Autonomy, Action, Accountability. The dimensions along which delegation to an agent actually escalates.
Evidence Pack
The reproducible artefact set behind an assurance claim. Determinism is the requirement: same inputs, same trace, same conclusion.
Alignment Readiness Score
A single tracked measure of whether a system's demonstrated behaviour matches the authority it has been granted.
Continuous TEVV
Assurance as an operating loop rather than a launch gate — because an agent's behaviour is not fixed at deployment.

Using the ladder

The ladder is only useful if a real use case can be placed on it and the placement changes what happens next. Three steps, in order — each one depends on the one before it.

  1. 01
    Place the use case at the mode it actually runs at

    Not the mode it was approved at, and not the mode the vendor describes. Ask what the system can reach, decide, and commit without a person. Where the answer surprises the owner, the placement has already earned its keep.

  2. 02
    Read the assurance floor for that mode

    The floor is cumulative — everything below the rung applies too. Gaps against the floor are the work plan, in priority order, and they are stated as controls rather than intentions so each one can be tested.

  3. 03
    Prove it, then keep proving it

    Test before deployment, monitor after, and collect both into the Evidence Pack. A use case that climbs a rung without re-clearing the floor is the failure mode the whole ladder exists to make visible.

Tiarne Hawkins — Scaling AI we can trust.