Agency Loop Map¶
Artifact — instantiates Agentic Control Loop Design
Lays the agent's full goal-to-feedback loop out as one connected diagram so a missing or broken coupling becomes visible at a glance.
The Agency Loop Map is a single diagram that draws the whole control loop as connected nodes — goal, world model, what the actor can observe, its action options, how it selects among them, whether it can actually execute, the feedback it gets, and the rule that updates the next move — and then marks each arrow between them as present, weak, or missing. Its defining move is that it works on the couplings, not the boxes: the failure it is built to catch is not an absent component but a live component that connects to nothing — observation that never reaches the choice, a choice that cannot be executed, an execution whose effects are never seen. It is a static picture that locates the break; it runs none of the links it draws. Where a dashboard shows whether a working loop is healthy, the map shows whether the loop exists at all.
Example¶
A regional food bank is missing its distribution targets and no one can say why. A coordinator sketches an Agency Loop Map for the delivery teams. Goal node: "get perishable donations to partner pantries before spoilage." Observation node: teams can see today's donation intake and their own truck's route. Action options: which pantries to visit, in what order, how much to drop. Selection: dispatchers phone each driver mid-route. Execution: trucks, drivers, fuel — all present.
Drawing the arrows exposes the break. Observation feeds selection fine, and selection feeds execution fine — but there is no arrow back from the pantries. Drivers drop pallets and leave; nobody records whether the food was actually used or landed at a pantry that was already full. The teams have observation, choice, and execution, but the effect-feedback arc is a dangling line. The map does not fix this — it makes the single missing coupling impossible to keep ignoring, and turns "why are we missing targets?" into "we need a pantry-side confirmation the choosing loop can see."
How it works¶
- Place the canonical nodes first. Draw the standard loop skeleton — goal, model, observation, options, selection, execution, feedback, update — before filling in this actor's specifics, so nothing is silently omitted.[n1]
- Trace one full lap. Follow a single real decision around the loop and draw only the arrows that genuinely carry information or authority. An arrow you cannot name a mechanism for is drawn dashed.
- Grade the three live arcs. The map's diagnostic weight is on the working middle: does what the actor observes actually reach the selection, and can the selected action actually be executed? These are the couplings that fail quietly.
- Circle the break, not the boxes. The output is the weakest coupling ringed on the diagram — the one repair that would let the lap close — handed to the mechanisms that populate the goal, model, and feedback nodes.
Tuning parameters¶
- Node granularity — one loop for the whole actor versus a sub-loop per task. Finer maps catch a break that hides inside an averaged loop but multiply diagrams; coarser maps stay legible but blur where exactly the arc snaps.
- Arc-evidence bar — how much proof an arrow needs before it is drawn solid. A strict bar exposes couplings people assume exist but cannot demonstrate; a loose bar draws an optimistic loop that looks whole and is not.
- Actor scope — mapping a person, a team, or a software agent. The node labels change (permission vs. capability, judgment vs. policy) and so does what counts as a broken arc.
- Refresh trigger — a one-time diagnostic versus a redraw whenever roles, tools, or authority shift. Redrawing keeps the map honest as the loop changes but adds upkeep.
When it helps, and when it misleads¶
Its strength is cheap, shared visibility: on one page, a whole room can see that the loop everyone assumed was closed has a dangling arc, and can argue about the one coupling that matters instead of trading blame. It is the survey the rest of the design builds on, and it is neutral about who is at fault — a missing arrow is a structural fact, not a person's failing.
Its failure mode is the map–territory error: a tidy diagram invites the belief that a drawn arrow is a working one, so a loop that is whole on paper and broken in practice passes inspection.[n1] It also flatters completeness — every box filled reads as health even when each link is thin. The guarding discipline is the arc-evidence bar run honestly: an arrow is solid only when someone can point to the mechanism that carries it, and the map is redrawn, not defended, when reality diverges.
How it implements the components¶
The map realizes the archetype's structural-diagnostic slice — the couplings it can render and grade, not the ones other mechanisms operate:
observation_channel— the map draws exactly what the actor can and cannot see, exposing observation that exists but never reaches a decision.selection_policy— it renders the node that turns goals and beliefs into a chosen action, and grades whether observation actually feeds it.execution_capability— it shows whether a selected action has the resources, permission, and tools to actually be carried out, catching choices that die at execution.
It renders the goal, model, and feedback nodes but populates none of them: represented_goal is confirmed by Briefback or Intent Confirmation, world_model is documented by Model Assumption Register, and the effect_feedback_loop is instrumented by Agency Health Dashboard. The map shows whether each link exists; the dashboard shows whether a working link is healthy.
Related¶
- Instantiates: Agentic Control Loop Design — the map is the first-pass diagnostic that finds which coupling the rest of the design must repair.
- Sibling mechanisms: Agency Health Dashboard · Model Assumption Register · After-Action Learning Cycle · Briefback or Intent Confirmation · Safe Action Menu · Graduated Autonomy Ramp · Decision-Rights Matrix · Controllability Mapping Checklist
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: The mechanism lays the agent's full goal-to-feedback loop out as one connected diagram so a missing or broken coupling becomes visible at a glance, so its operative form is a static or prospective information artifact.
Independent corroboration: The frozen evidence defines Agency Loop Map as 'Lays the agent's full goal-to-feedback loop out as one connected diagram so a missing or broken coupling becomes visible at a glance', so its operative form is Representation, Specification & Plan.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Systems Thinking & Cybernetics
Origin pattern: Convergent development
Present-day reach: Universal
Rationale: Cybernetic control diagrams connect goals, observation, internal model, selection, execution, effect feedback, and update, locating failures in missing couplings.
Related originating lineages:
- Engineering & Design — Functional block diagrams and interface analysis contribute the artifact form.
- Human-Computer Interaction — Visibility of action and feedback couplings contributes human-agent applications.
- Military & Strategic Studies — Boyd's OODA loop supplies a compact observe-orient-decide-act cycle and emphasis on speed and integrity of connections.
- Organizational & Management Science — Decision rights and broken organizational handoffs provide a major social-system application.
- Psychology — Models of agency and self-regulation contribute intention, perceived action options, efficacy, attribution, and behavioral feedback.
Review resolution: Cybernetic control diagrams are primary, while engineering block diagrams, HCI feedback, OODA, organizational decision rights, and psychological self-regulation independently develop materially similar couplings. Universal reach describes applicability without multiplying origins.
Attribution caveat: The full coupling map is cybernetic; OODA is its clearest named agentic operational cousin.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Boyd's OODA loop (Observe–Orient–Decide–Act) is the canonical minimal control cycle; the map checks a real actor against a skeleton like it. The map–territory caution — Korzybski's "the map is not the territory" — is the standing hazard: a drawn loop is a claim about the real one, not proof of it. ↩a ↩b