Tiarne Hawkins

Scoping agentic AI

The Four A’s

Four questions that define an agent’s operating envelope. Answer them before deployment, or find out what the answers were afterwards.

Most agentic AI reviews start with the model. That is the wrong end of the problem. A model that hallucinates on a read-only research task is an annoyance; the same model wired to a payments API with no escalation threshold is an incident.

What separates the two is not intelligence. It is envelope — the set of things the system is permitted to reach, decide, change, and answer for. This framework scopes every agentic system across four dimensions, in the order a request actually moves through them.

Gate 01

Access

What can this system reach?

Access is the reachability question, and it is almost always answered by accident. An agent gets a service account because that was the fastest path through the ticket queue, and inherits entitlements no human on the team holds. Nobody granted that. Nobody revoked it either.

Scoping Access means naming every system, dataset, tool and credential the agent can touch — including the ones it reaches transitively, through a tool that calls another tool. Reachability is not what the design document says. It is what the runtime permits.

Controls

Entitlement scoping per agent identity, a governed tool registry, data classification enforced at the boundary rather than in the prompt.

Failure mode

Privilege inheritance. The agent can read a table nobody remembers connecting, and the exposure only surfaces in the breach review.

Evidence

A reachability map, generated from the runtime and diffed on every release.

Gate 02

Autonomy

How far does it get to go on its own?

Autonomy is the axis the M0–M6 maturity model measures. It runs from a system that executes a fixed script to one that sets its own sub-goals and calls its own tools across a multi-step task. Most enterprises are further up that ladder than their governance assumes, because autonomy is rarely granted — it accrues. A checkpoint gets removed to hit a latency target, and the review that would have caught it was scoped to the model, not the workflow.

Scoping Autonomy means setting the escalation threshold explicitly, in terms the business already uses: value at risk, reversibility, confidence, and precedent. Then measuring how often the threshold actually fires.

Controls

Declared maturity level per system, escalation thresholds tied to value and reversibility, human-in-the-loop checkpoints that are tested rather than assumed.

Failure mode

Silent escalation drift. The override rate falls to near zero and gets read as accuracy, when it is really a reviewer approving whatever appears.

Evidence

Escalation and override logs, with the threshold that triggered each one.

Gate 03

Action

What can it change in the world?

Access is what the agent can see. Action is what it can move. The two get conflated constantly, and that conflation is where blast radius hides — a system described as read-only in every design review, holding one write-capable tool because the SDK shipped it in the same client.

Scoping Action means classifying every effect the agent can produce into reversibility tiers, and putting hard limits on the tiers that matter: rate, value, counterparty, time window. Reversible actions can be tested in production. Irreversible ones have to be proven before they get there.

Controls

An action inventory with reversibility tiers, blast-radius caps on rate and value, irreversible effects gated behind explicit approval.

Failure mode

An unlisted write path. The agent's intended behaviour is read-only; its capability is not, and no test exercised the gap.

Evidence

Pre-deployment test results against the declared action set, including the ones that should have failed and did.

Gate 04

Accountability

Who answers when it is wrong?

Every regulated firm can produce logs. Very few can reconstruct a decision. There is a difference: a log shows what the model returned, and a reconstruction shows the inputs it saw, the version that was serving, the tools it called, the threshold it cleared, and the person whose name sits on the outcome.

Accountability is the gate that converts an agent from a system you operate into a system you can defend. It is also the one that has to be designed in at the start, because a decision record cannot be assembled retrospectively from telemetry that was never captured.

Controls

A named accountable owner per system, immutable decision records, model and prompt versioning bound to every run.

Failure mode

Orphaned ownership. The agent outlives the team that built it, and the answer to “who owns this” is a distribution list.

Evidence

An Evidence Pack: the reconstructable chain, assembled continuously rather than at examination time.

Same use case, two envelopes

The Four A’s are not a maturity score on their own. They are the profile a maturity level describes. Two dispute-intake agents in the same retail bank, doing recognisably the same job.

Drafting copilotResolution agent
AccessCase notes, read-only. No customer PII leaves the boundary.Case notes, payments API, customer comms, prior-dispute history.
AutonomyProposes only. A handler sends, every time.Resolves below a set value threshold; escalates above it, and on any novel pattern.
ActionWrites nothing. The draft is an artefact, not an effect.Issues provisional credit and sends customer comms. Reversible within a fixed window; irreversible ledger entries blocked.
AccountabilityNamed handler on every send. Draft retained with the case.Named owner, immutable decision record per resolution, sampled human review with a published sampling rate.

Same model, potentially. Entirely different assurance obligation. The second one is not riskier because it is smarter — it is riskier because three of its four gates are open wider.

Where the Four A’s live

The framework is only useful if each gate has somewhere to be declared, tested, watched and proven.

Use Cases & Alignment

Where the four gates are declared, per system, in a registry the risk function can actually read.

AI Range

Where the declaration gets tested before deployment. Claimed envelope against actual behaviour.

Peregrine

Where drift shows up after deployment. Gates widen quietly; this is what notices.

Evidence Pack

Where the record lands, in the form an examiner will ask for.

Trust in an agentic system is an engineered output, not a claim made about a model. The Four A’s are where the engineering starts: four answers, written down, tested, and kept current as the system changes.

Scope your first system — discuss an engagement
Tiarne Hawkins — Scaling AI we can trustAccess · Autonomy · Action · Accountability