The agentic model.

In brief

The model is deliberately compact: one quartet of attributes that any decision must hold in alignment, tested decision by decision rather than agent by agent, one definition of an agent that spans people, teams, systems, and AI, and a short list of process elements that connect agents and outcomes. Most useful findings live in the interactions between these rather than in any single one, and most of the value lies in identifying a misalignment precisely enough that the move to fix it becomes obvious.

The model is not tied to any one sector. It describes any process in which agents make and act on decisions, whatever the industry. Thinkering's own depth is in healthcare commercial and operational work: funding model design, claims integrity, provider network strategy, contract negotiation. Although the domain areas and examples for each engagement are specific, the frameworks that we apply are industry-agnostic.

Thinkering's agentic model is a working draft. None of this is presented as new theory, it is a synthesis of established work, assembled into a working instrument and the foundations are shared here in the form that we currently use.

The diagnostic quartet

Four attributes that must align at each decision an organisation makes.

• Knowledge. What is known about the situation the decision concerns.
• Authority. Who is permitted to decide and within what bounds.
• Accountability. Who answers for the outcome.
• Capacity. The available action that the system can take to deliver on the decision.

Misalignment between the four is the failure mode that we diagnose and redesign. Knowledge without authority becomes recommendation that is never acted on. Authority without accountability becomes drift. Capacity without knowledge becomes throughput in the wrong direction.

Draws on

Jensen and Meckling (1992); Brickley, Smith and Zimmerman (1997); Bossert (1998).

The agent

An agent is anything a process hands a piece of work to and depends on the result of without reproducing the working: a person, a team, a system, or an AI. Four kinds of agent fall under the same definition.

• People. Individual practitioners.
• Teams and committees. Multi-person actors with discretion held collectively.
• Systems. Rules engines, workflow systems, codified procedures.
• AI. Models with discretion over inputs, scoring, or output.

The kind of agent matters in practice. It does not matter in description: the same seven attributes apply across all four, and a deterministic rules engine and a senior clinician are both agents, differing in the values those attributes take, discretion among them.

Draws on

Russell and Norvig (1995); Ashby (1956); Beer (1979); Dennett (1987); List and Pettit (2011); Lipsky (1980); Latour (2005); van der Aalst et al. (2023).

The seven attributes

Each agent, regardless of kind, is characterised by seven attributes.

• Scope of action. What the agent can do.
• Discretion. Within scope, what the agent decides for itself.
• Inputs and outputs. What triggers the agent, what it produces, the hand-offs it owns.
• Capabilities. The kinds of work the agent can actually do well.
• Accountability. To whom the agent answers and for what.
• Supervision. How the agent's action is observed and corrected.
• Capacity. How much the agent can do, and where it degrades.

The seven attributes map onto the quartet:

Most diagnostic findings live in the interactions between attributes (ie. scope without capability, discretion without supervision, capacity without intake control) rather than in any single attribute alone.

Draws on

Wooldridge, Jennings and Kinny (2000); Lipsky (1980); Jaques (1976); Parasuraman, Sheridan and Wickens (2000); Hollnagel and Woods (2005).

The four process elements

A process is the sequence of agent actions connected by hand-offs. Four elements bound it.

• Intent. What the process exists to produce (output intent) and what that output is meant to enable (outcome intent).
• Hand-offs. Where one agent's output becomes another's trigger. A clean hand-off transfers state, accountability, and trigger conditions.
• Feedback. How the system learns whether intent was met, at the agent level and at the system level.
• Constraints. What the process is bounded by externally: policy, evidence, regulation, contractual terms, resource limits.

Draws on

IDEF0 (FIPS PUB 183); Checkland (1981); Donabedian (1966).

The decision

A decision is the point in a process where an agent turns an input into an output the process relies on. It is where the quartet is tested: knowledge, authority, accountability and capacity have to align for the particular call being made, not for the agent in general. The same agent can hold one decision where the four align and another where they do not.

An agent can be opened into the decisions it holds, but the unit that resolution stops at is the decision class, not the single call: a recurring decision identified by its intent, action, consequence and required discretion, with the cases that share these read as one. An agent who makes dozens of calls a day holds only a handful of classes. Two calls belong to the same class while they do not differ in a way that changes the diagnosis; a class is split when a subset needs different discretion or carries a different consequence, since those are what change who should hold it. A rules engine opened the same way resolves to a few rule-classes, not one decision for every rule it holds.

Beyond the quartet, a decision class carries four things of its own, the same four by which decisions are grouped into classes.

• Intent. What the decision produces and what that output is meant to enable, in the same two layers as the process.
• Action. The form the output takes: stop, warn, refer, apply, or pass. A refer is the decision routing a case onward rather than resolving it.
• Consequence. The cost of deciding wrongly, which is what supervision should be calibrated against rather than historical habit.
• Required discretion. The judgement the decision demands, read against the discretion the holding agent supplies. Where a decision needs judgement its agent cannot give and there is no path to escalate, discretion is being denied where the work needs it.

Draws on

Donabedian (1966); IDEF0 (FIPS PUB 183); Beer (1979); Sheridan and Verplank (1978); Jaques (1976, 1998); Simon (1960); Decision Model and Notation (OMG 2015).

Applying the model

The framework is applied in six steps. Each produces an artefact.

1. Scope the process. Name the process, its intent at both layers, what counts as completion.
2. List the agents. Identify every actor across all four kinds. Characterise each by the seven attributes.
3. Map the flow. Sequence the agent actions and hand-offs. Note feedback loops and constraints.
4. Diagnose each decision. For each decision in the process, locate knowledge, authority, accountability, and capacity. Open an agent into its decisions where they differ in kind, and read the discretion each decision requires against what its agent supplies. Name where the four are misaligned.
5. Propose the moves, when in scope. Propose moves at the agent level: introduce or remove agents, reshape scope or discretion, redesign hand-offs, recalibrate supervision.
6. Make the capacity case. State what the work shows about capacity in this operation and which levers the recommended moves use.

The sequence is iterative, not linear. The agent register often reveals that the process was scoped too narrowly or too broadly, and the diagnostic step often reveals that an agent was missed. The artefacts are revised as the understanding develops.